Discussion:
benefits of ack filtering
(too old to reply)
Dave Taht
2017-11-28 21:48:31 UTC
Permalink
Raw Message
Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.

The benefits were pretty impressive at a 50x1 Down/Up ratio:

http://blog.cerowrt.org/post/ack_filtering/

And quite noticeable at 16x1 ratios as well.

I'd rather like to have a compelling list of reasons why not to do
this! And ways to do it better, if not. The relevant code is hovering
at:

https://github.com/dtaht/sch_cake/blob/cobalt/sch_cake.c#L902
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Mikael Abrahamsson
2017-11-29 06:09:17 UTC
Permalink
Raw Message
Post by Dave Taht
Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.
http://blog.cerowrt.org/post/ack_filtering/
And quite noticeable at 16x1 ratios as well.
I'd rather like to have a compelling list of reasons why not to do
this! And ways to do it better, if not. The relevant code is hovering
https://github.com/dtaht/sch_cake/blob/cobalt/sch_cake.c#L902
Your post is already quite comprehensive when it comes to downsides.

The better solution would of course be to have the TCP peeps change the
way TCP works so that it sends fewer ACKs. I don't want middle boxes
making "smart" decisions when the proper solution is for both end TCP
speakers to do less work by sending fewer ACKs. In the TCP implementations
I tcpdump regularily, it seems they send one ACK per 2 downstream packets.

At 1 gigabit/s that's in the order of 35k pps of ACKs (100 megabyte/s
divided by 1440 divided by 2). That's in my opinion completely ludicrous
rate of ACKs for no good reason.

I don't know what the formula should be, but it sounds like the ACK
sending ratio should be influenced by how many in-flight ACKs there might
be. Is there any reason to have more than 100 ACKs in flight at any given
time? 500? 1000?

My DOCSIS connection (inferred through observation) seems to run on 1ms
upstream time slots, and my modem will delete contigous ACKs at 16 or 32
ACK intervals, ending up running at typically 1-2 ACKs per 1ms time slot.
This cuts down the ACK rate when I do 250 megabit/s downloads from 5-8
megabit/s to 400 kilobit/s of used upstream bw.

Since this ACK reduction is done on probably hundreds of millions of
fixed-line subscriber lines today, what arguments do designers of TCP have
to keep sending one ACK per 2 received TCP packets?
--
Mikael Abrahamsson email: ***@swm.pp.se
Sebastian Moeller
2017-11-29 09:34:08 UTC
Permalink
Raw Message
Well, ACK filtering/thinning is a simple trade-off: redundancy versus bandwidth. Since the RFCs say a receiver should acknoledge every second full MSS I think the decision whether to filter or not should be kept to the enduser and not some misguided middle boxes; if a DOCSIS ISP wants to secure precious upstream bandwidth they should at least re-synthesize the filtered ACKs after passing their upstream bottleneck IMHO. This is not reasonable network management in my irrelevant opinion unless actively opted-in by the user. Or put differently the real fix for DOCSIS ISPs is to simply not sell internet connections with asymmetries that make it impossible to saturate the link with TCP traffic without heroic measures like ack filtering.
So I am all for cake learning to do that, but I am 100% against recommending using it unless one is "blessed" with a clue-less ISP that has problems calculating the maximal permissible Up/Down asymmetry for TCP...
BTW, I believe older TCPs used the reception of an ACK and not the acknowledged byte increment for widening their send/congestion windows, ack filtering should make slow start behave more sluggish for such hosts. As far as I can tell linux recently learned to deal with this fact as GRO in essence will also make the receiver ACK more rarely (once every 2 super-packets), so linux I think now evaluates the number of acknoledged bytes. But I have no idea about windows or BSD tcp implementations.

Best Regards
Post by Mikael Abrahamsson
Post by Dave Taht
Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.
http://blog.cerowrt.org/post/ack_filtering/
And quite noticeable at 16x1 ratios as well.
I'd rather like to have a compelling list of reasons why not to do
this! And ways to do it better, if not. The relevant code is hovering
https://github.com/dtaht/sch_cake/blob/cobalt/sch_cake.c#L902
Your post is already quite comprehensive when it comes to downsides.
The better solution would of course be to have the TCP peeps change the way TCP works so that it sends fewer ACKs. I don't want middle boxes making "smart" decisions when the proper solution is for both end TCP speakers to do less work by sending fewer ACKs. In the TCP implementations I tcpdump regularily, it seems they send one ACK per 2 downstream packets.
At 1 gigabit/s that's in the order of 35k pps of ACKs (100 megabyte/s divided by 1440 divided by 2). That's in my opinion completely ludicrous rate of ACKs for no good reason.
I don't know what the formula should be, but it sounds like the ACK sending ratio should be influenced by how many in-flight ACKs there might be. Is there any reason to have more than 100 ACKs in flight at any given time? 500? 1000?
My DOCSIS connection (inferred through observation) seems to run on 1ms upstream time slots, and my modem will delete contigous ACKs at 16 or 32 ACK intervals, ending up running at typically 1-2 ACKs per 1ms time slot. This cuts down the ACK rate when I do 250 megabit/s downloads from 5-8 megabit/s to 400 kilobit/s of used upstream bw.
Since this ACK reduction is done on probably hundreds of millions of fixed-line subscriber lines today, what arguments do designers of TCP have to keep sending one ACK per 2 received TCP packets?
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Mikael Abrahamsson
2017-11-29 12:49:24 UTC
Permalink
Raw Message
Post by Sebastian Moeller
Well, ACK filtering/thinning is a simple trade-off: redundancy versus
bandwidth. Since the RFCs say a receiver should acknoledge every second
full MSS I think the decision whether to filter or not should be kept to
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP
transfer?

Sounds like a lot of useless interrupts and handling by the stack, apart
from offloading it to the NIC to do a lot of handling of these mostly
useless packets so the CPU doesn't have to do it.

Why isn't 1kPPS of ACKs sufficient for most usecases?
--
Mikael Abrahamsson email: ***@swm.pp.se
Luca Muscariello
2017-11-29 13:13:05 UTC
Permalink
Raw Message
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1

It would be interesting to know what is the minimum ACK rate to achieve
full utilisation.
Or the how the downlink rate depends on the uplink ACK rate.
I'm sure I've seen this dependency in some old paper.
Post by Sebastian Moeller
Well, ACK filtering/thinning is a simple trade-off: redundancy versus
Post by Sebastian Moeller
bandwidth. Since the RFCs say a receiver should acknoledge every second
full MSS I think the decision whether to filter or not should be kept to
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Sounds like a lot of useless interrupts and handling by the stack, apart
from offloading it to the NIC to do a lot of handling of these mostly
useless packets so the CPU doesn't have to do it.
Why isn't 1kPPS of ACKs sufficient for most usecases?
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Mikael Abrahamsson
2017-11-29 14:31:30 UTC
Permalink
Raw Message
Post by Luca Muscariello
Post by Mikael Abrahamsson
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1
RFC3449 is all about middleboxes doing things.

I wanted to understand why TCP implementations find it necessary to send
one ACK per 2xMSS at really high PPS. Especially when NIC offloads and
middleboxes frequently strip out this information anyway so it never
reaches the IP stack (right?).
--
Mikael Abrahamsson email: ***@swm.pp.se
Jonathan Morton
2017-11-29 14:36:03 UTC
Permalink
Raw Message
There is an RFC about emitting fewer acks in response to upstream
congestion or some other cue; AckCC might be the name.

- Jonathan Morton
Andrés Arcia-Moret
2017-11-29 15:24:55 UTC
Permalink
Raw Message
RFC 5690
and
http://www.saber.ula.ve/bitstream/123456789/30345/1/thesis-francais-english.pdf <http://www.saber.ula.ve/bitstream/123456789/30345/1/thesis-francais-english.pdf>
Ch 4 and 5

regards

a.
There is an RFC about emitting fewer acks in response to upstream congestion or some other cue; AckCC might be the name.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Luca Muscariello
2017-11-29 15:53:43 UTC
Permalink
Raw Message
Post by Mikael Abrahamsson
Why does it say to do this? What benefit is there to either end system to
Post by Luca Muscariello
Post by Sebastian Moeller
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1
RFC3449 is all about middleboxes doing things.
I wanted to understand why TCP implementations find it necessary to send
one ACK per 2xMSS at really high PPS. Especially when NIC offloads and
middleboxes frequently strip out this information anyway so it never
reaches the IP stack (right?).
I would say because it is complex to guess at which PPS to work. You would
need an adaptation mechanism. Need also to change the client and the server
sides. The AckCC Jonathan has mentioned
might be a solution to that.
Probably an ACK pacer in the end host, out of the TCP stack, doing Ack
filtering and decimation can be simpler to implement than the proper
adaptation mechanism in TCP.
Maybe inside sch_fq it would be doable. Maybe not.
Jonathan Morton
2017-11-29 16:13:47 UTC
Permalink
Raw Message
Given an RTT estimate and knowledge of the congestion window, the AckCC
option could be used to target a handful of acks (maybe 4 to 10) per RTT.
As usual, extra acks would be sent when loss is suspected, on ECN events,
and when the push flag is set.

That would be perfectly sufficient.

- Jonathan Morton
Michael Welzl
2017-11-30 07:03:40 UTC
Permalink
Raw Message
Hi Bloaters,

I’d like to give offer some information and thoughts on AckCC, at the bottom of this email.
Post by Mikael Abrahamsson
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1 <https://tools.ietf.org/html/rfc3449#section-5.2.1>
RFC3449 is all about middleboxes doing things.
I wanted to understand why TCP implementations find it necessary to send one ACK per 2xMSS at really high PPS. Especially when NIC offloads and middleboxes frequently strip out this information anyway so it never reaches the IP stack (right?).
I would say because it is complex to guess at which PPS to work. You would need an adaptation mechanism. Need also to change the client and the server sides. The AckCC Jonathan has mentioned
might be a solution to that.
Probably an ACK pacer in the end host, out of the TCP stack, doing Ack filtering and decimation can be simpler to implement than the proper adaptation mechanism in TCP.
Maybe inside sch_fq it would be doable. Maybe not.
I’m adding the response from Jonathan Morton here to make this more self-contained:
***
Given an RTT estimate and knowledge of the congestion window, the AckCC option could be used to target a handful of acks (maybe 4 to 10) per RTT. As usual, extra acks would be sent when loss is suspected, on ECN events, and when the push flag is set.
That would be perfectly sufficient.

- Jonathan Morton

***

A few years ago, David Ros, whom I’m adding in cc, one of the original authors of RFC 5690 did a sabbatical with me at the University of Oslo. As part of that, we advised a master student to carry out tests with AckCC, and analyze the RFC to understand how it would have to change if we were to proceed to Proposed Standard. The result of his investigation is here:
http://heim.ifi.uio.no/michawe/teaching/dipls/marius-olsen/mastersthesis-mariusno.pdf <http://heim.ifi.uio.no/michawe/teaching/dipls/marius-olsen/mastersthesis-mariusno.pdf>
and his code is here: http://folk.uio.no/mariusno/master/ <http://folk.uio.no/mariusno/master/>

Now, after finishing the thesis, when it came to writing a paper about it, we got stuck in the discussion of “how are we going to explain that this is really necessary?”
- we didn’t want to submit a “solution searching for a problem” paper and didn’t want to get rejected for not having shown that the problem truly exists. Searching for this a little in the academic world (papers) gave us no result, at least back then.

Interestingly, at IETF 98, not so long ago, Ingemar Johansson explained to folks at TSVWG that the problem IS real:
https://datatracker.ietf.org/meeting/98/materials/slides-98-tsvwg-sessb-7-transport-protocol-feedback-overhead-issues-and-solutions/ <https://datatracker.ietf.org/meeting/98/materials/slides-98-tsvwg-sessb-7-transport-protocol-feedback-overhead-issues-and-solutions/>

So, let me now try to answer “why is TCP not doing that?”.
- First, AFAIK, AckCC isn’t implemented anywhere (except that we have this old patch - please feel free to download, adapt, and play with it !!)
- Second, if someone was to update TCP to support this, a bit more than simple statements about the amount of traffic being large would be good IMO - I mean, some convincing proof that the large number of ACKs *really* is a problem.
- Third, once this is implemented and deployed and found to be beneficial, it would be useful to follow up in the IETF and update RFC 5690.

Since nobody seems to be doing any of these things, nothing changes. But consider this: I see folks from Google doing a lot of TCP updates in the IETF for which they themselves appear to have an immediate need. Given the heterogeneity and magnitude of traffic produced by Google, if they don’t see a pressing need for it, I suspect that, indeed, the problem might not be so real after all?!

Also, a word of caution. In this thread, there seems to be general agreement that TCP sends way too many ACKs, and that reducing that number would be fantastic.
I’m not so convinced. Okay, even if TCP isn’t that ACK-clocked anymore in Linux: 1) there isn’t only Linux in this world, 2) ACKs are still quite important in Fast Recovery, 3) BBR might not need to clock out ACKs, but it measures their incoming rate. For another example, consider a non-BBR sender in slow start: without ABC, missing ACKs would let it grow its cwnd too cautiously. Thanks to ABC, this can be done more aggressively - but ABC recommends a limit on how quickly to “jump” in the rate in response to a single ACK, for good reason - to avoid producing even heavier bursts. But with this limit, again, the TCP sender is unnecessarily cautious in Slow Start just because it misses ACKs.

My point is: the ACKs ARE the feedback that TCP works on; when you take them away, TCP becomes “blind”, and whatever improvement is made to TCP will have to be developed on that basis.

I’m not saying that 1 ACK for every two packets is really necessary
 but unless there’s hard proof that this really is a problem, I’d caution against a “downward spiral” here: the level of asymmetry offered to users today is probably somehow related to the commonly seen TCP ACK rate - so if TCP starts to reduce the ACK rate, folks may decide to make links even more asymmetric, etc. etc. 
 I’m not sure this is a good direction.

Just some thoughts, and some context.

Cheers,
Michael
Dave Taht
2017-11-30 07:24:27 UTC
Permalink
Raw Message
Post by Michael Welzl
Hi Bloaters,
I’d like to give offer some information and thoughts on AckCC, at the bottom
of this email.
Post by Mikael Abrahamsson
Post by Luca Muscariello
Post by Mikael Abrahamsson
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1
RFC3449 is all about middleboxes doing things.
I wanted to understand why TCP implementations find it necessary to send
one ACK per 2xMSS at really high PPS. Especially when NIC offloads and
middleboxes frequently strip out this information anyway so it never reaches
the IP stack (right?).
I would say because it is complex to guess at which PPS to work. You would
need an adaptation mechanism. Need also to change the client and the server
sides. The AckCC Jonathan has mentioned
might be a solution to that.
Probably an ACK pacer in the end host, out of the TCP stack, doing Ack
filtering and decimation can be simpler to implement than the proper
adaptation mechanism in TCP.
Maybe inside sch_fq it would be doable. Maybe not.
I’m adding the response from Jonathan Morton here to make this more
***
Given an RTT estimate and knowledge of the congestion window, the AckCC
option could be used to target a handful of acks (maybe 4 to 10) per RTT.
As usual, extra acks would be sent when loss is suspected, on ECN events,
and when the push flag is set.
That would be perfectly sufficient.
- Jonathan Morton
***
A few years ago, David Ros, whom I’m adding in cc, one of the original
authors of RFC 5690 did a sabbatical with me at the University of Oslo. As
part of that, we advised a master student to carry out tests with AckCC, and
analyze the RFC to understand how it would have to change if we were to
http://heim.ifi.uio.no/michawe/teaching/dipls/marius-olsen/mastersthesis-mariusno.pdf
and his code is here: http://folk.uio.no/mariusno/master/
Now, after finishing the thesis, when it came to writing a paper about it,
we got stuck in the discussion of “how are we going to explain that this is
really necessary?”
- we didn’t want to submit a “solution searching for a problem” paper and
didn’t want to get rejected for not having shown that the problem truly
exists. Searching for this a little in the academic world (papers) gave us
no result, at least back then.
Interestingly, at IETF 98, not so long ago, Ingemar Johansson explained to
https://datatracker.ietf.org/meeting/98/materials/slides-98-tsvwg-sessb-7-transport-protocol-feedback-overhead-issues-and-solutions/
So, let me now try to answer “why is TCP not doing that?”.
- First, AFAIK, AckCC isn’t implemented anywhere (except that we have this
old patch - please feel free to download, adapt, and play with it !!)
- Second, if someone was to update TCP to support this, a bit more than
simple statements about the amount of traffic being large would be good IMO
- I mean, some convincing proof that the large number of ACKs *really* is a
problem.
- Third, once this is implemented and deployed and found to be beneficial,
it would be useful to follow up in the IETF and update RFC 5690.
Since nobody seems to be doing any of these things, nothing changes. But
consider this: I see folks from Google doing a lot of TCP updates in the
IETF for which they themselves appear to have an immediate need. Given the
heterogeneity and magnitude of traffic produced by Google, if they don’t see
a pressing need for it, I suspect that, indeed, the problem might not be so
real after all?!
Also, a word of caution. In this thread, there seems to be general agreement
that TCP sends way too many ACKs, and that reducing that number would be
fantastic.
I’m not so convinced. Okay, even if TCP isn’t that ACK-clocked anymore in
Linux: 1) there isn’t only Linux in this world, 2) ACKs are still quite
important in Fast Recovery, 3) BBR might not need to clock out ACKs, but it
measures their incoming rate. For another example, consider a non-BBR
sender in slow start: without ABC, missing ACKs would let it grow its cwnd
too cautiously. Thanks to ABC, this can be done more aggressively - but ABC
recommends a limit on how quickly to “jump” in the rate in response to a
single ACK, for good reason - to avoid producing even heavier bursts. But
with this limit, again, the TCP sender is unnecessarily cautious in Slow
Start just because it misses ACKs.
My answer to questions like this that are difficult reason about... is
to run the experiment.

Trying out BBR in the testbeds we have setup would be straightforward,
although rrul_be (which is what we have the MOS results for) is not
the best test for BBR's behaviors.

Maybe more of a staircase test would be better.

(note we're also looking at sfq and pfifo as references)
Post by Michael Welzl
My point is: the ACKs ARE the feedback that TCP works on; when you take them
away, TCP becomes “blind”, and whatever improvement is made to TCP will have
to be developed on that basis.
I’m not saying that 1 ACK for every two packets is really necessary… but
unless there’s hard proof that this really is a problem, I’d caution against
a “downward spiral” here: the level of asymmetry offered to users today is
probably somehow related to the commonly seen TCP ACK rate - so if TCP
starts to reduce the ACK rate, folks may decide to make links even more
asymmetric, etc. etc. … I’m not sure this is a good direction.
Just some thoughts, and some context.
Cheers,
Michael
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Dave Taht
2017-11-30 07:45:12 UTC
Permalink
Raw Message
Post by Michael Welzl
Hi Bloaters,
I’d like to give offer some information and thoughts on AckCC, at the bottom
of this email.
Post by Mikael Abrahamsson
Post by Luca Muscariello
Post by Mikael Abrahamsson
Why does it say to do this? What benefit is there to either end system to
send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Did you check RFC 3449 ?
https://tools.ietf.org/html/rfc3449#section-5.2.1
RFC3449 is all about middleboxes doing things.
I wanted to understand why TCP implementations find it necessary to send
one ACK per 2xMSS at really high PPS. Especially when NIC offloads and
middleboxes frequently strip out this information anyway so it never reaches
the IP stack (right?).
I would say because it is complex to guess at which PPS to work. You would
need an adaptation mechanism. Need also to change the client and the server
sides. The AckCC Jonathan has mentioned
might be a solution to that.
Probably an ACK pacer in the end host, out of the TCP stack, doing Ack
filtering and decimation can be simpler to implement than the proper
adaptation mechanism in TCP.
Maybe inside sch_fq it would be doable. Maybe not.
I’m adding the response from Jonathan Morton here to make this more
***
Given an RTT estimate and knowledge of the congestion window, the AckCC
option could be used to target a handful of acks (maybe 4 to 10) per RTT.
As usual, extra acks would be sent when loss is suspected, on ECN events,
and when the push flag is set.
That would be perfectly sufficient.
- Jonathan Morton
***
A few years ago, David Ros, whom I’m adding in cc, one of the original
authors of RFC 5690 did a sabbatical with me at the University of Oslo. As
part of that, we advised a master student to carry out tests with AckCC, and
analyze the RFC to understand how it would have to change if we were to
http://heim.ifi.uio.no/michawe/teaching/dipls/marius-olsen/mastersthesis-mariusno.pdf
and his code is here: http://folk.uio.no/mariusno/master/
Now, after finishing the thesis, when it came to writing a paper about it,
we got stuck in the discussion of “how are we going to explain that this is
really necessary?”
- we didn’t want to submit a “solution searching for a problem” paper and
didn’t want to get rejected for not having shown that the problem truly
exists. Searching for this a little in the academic world (papers) gave us
no result, at least back then.
Interestingly, at IETF 98, not so long ago, Ingemar Johansson explained to
https://datatracker.ietf.org/meeting/98/materials/slides-98-tsvwg-sessb-7-transport-protocol-feedback-overhead-issues-and-solutions/
So, let me now try to answer “why is TCP not doing that?”.
- First, AFAIK, AckCC isn’t implemented anywhere (except that we have this
old patch - please feel free to download, adapt, and play with it !!)
- Second, if someone was to update TCP to support this, a bit more than
simple statements about the amount of traffic being large would be good IMO
- I mean, some convincing proof that the large number of ACKs *really* is a
problem.
- Third, once this is implemented and deployed and found to be beneficial,
it would be useful to follow up in the IETF and update RFC 5690.
Since nobody seems to be doing any of these things, nothing changes. But
consider this: I see folks from Google doing a lot of TCP updates in the
IETF for which they themselves appear to have an immediate need. Given the
heterogeneity and magnitude of traffic produced by Google, if they don’t see
a pressing need for it, I suspect that, indeed, the problem might not be so
real after all?!
Also, a word of caution. In this thread, there seems to be general agreement
that TCP sends way too many ACKs, and that reducing that number would be
fantastic.
I’m not so convinced. Okay, even if TCP isn’t that ACK-clocked anymore in
Linux: 1) there isn’t only Linux in this world,
Nor one Linux.
Post by Michael Welzl
2) ACKs are still quite
important in Fast Recovery,
If you are already achieving twice the rate, what does occasionally
losing fast recovery cost?
Post by Michael Welzl
3) BBR might not need to clock out ACKs, but it
measures their incoming rate.
if it collapses to a punctuated paced source, it could also notice
acks being lost, and extrapolate.
Post by Michael Welzl
For another example, consider a non-BBR
sender in slow start: without ABC, missing ACKs would let it grow its cwnd
too cautiously. Thanks to ABC, this can be done more aggressively - but ABC
recommends a limit on how quickly to “jump” in the rate in response to a
single ACK, for good reason - to avoid producing even heavier bursts. But
with this limit, again, the TCP sender is unnecessarily cautious in Slow
Start just because it misses ACKs.
I'm not a huge fan of slow start in IW10.

And: Strike "unnecessarily is", and substitute "may not be", as
Loading Image...
seems to show. The ack-filter result shows one flow growing rapidly,
and three others not.
Post by Michael Welzl
My point is: the ACKs ARE the feedback that TCP works on; when you take them
away, TCP becomes “blind”, and whatever improvement is made to TCP will have
to be developed on that basis.
I’m not saying that 1 ACK for every two packets is really necessary… but
unless there’s hard proof that this really is a problem, I’d caution against
a “downward spiral” here: the level of asymmetry offered to users today is
probably somehow related to the commonly seen TCP ACK rate - so if TCP
starts to reduce the ACK rate, folks may decide to make links even more
asymmetric, etc. etc. … I’m not sure this is a good direction.
Just some thoughts, and some context.
Cheers,
Michael
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Jonathan Morton
2017-11-30 07:48:24 UTC
Permalink
Raw Message
I do see your arguments. Let it be known that I didn't initiate the
ack-filter in Cake, though it does seem to work quite well.

With respect to BBR, I don't think it depends strongly on the return rate
of acks in themselves, but rather on the rate of sequence number advance
that they indicate. For this purpose, having the receiver emit sparser but
still regularly spaced acks would be better than having some middlebox
delete some less-predictable subset of them. So I think BBR could be a
good testbed for AckCC implementation, especially as it is inherently paced
and thus doesn't suffer from burstiness as a conventional ack-clocked TCP
might.

The real trouble with AckCC is that it requires implementation on the
client as well as the server. That's most likely why Google hasn't tried
it yet; there are no receivers in the wild that would give them valid data
on its effectiveness. Adding support in Linux would help here, but aside
from Android devices, Linux is only a relatively small proportion of
Google's client traffic - and Android devices are slow to pick up new
kernel features if they can't immediately turn it into a consumer-friendly
bullet point.

Meanwhile we have highly asymmetric last-mile links (10:1 is typical, 50:1
is occasionally seen), where a large fraction of upload bandwidth is
occupied by acks in order to fully utilise the download bandwidth in TCP.
Any concurrent upload flows have to compete with that dense ack flow, which
in various schemes is unfair to either the upload or the download
throughput.

That is a problem as soon as you have multiple users on the same link, eg.
a family household at the weekend. Thinning out those acks in response to
uplink congestion is a solution. Maybe not the best possible solution, but
a deployable one that works.

- Jonathan Morton
Luca Muscariello
2017-11-30 08:00:26 UTC
Permalink
Raw Message
Agree and think this is a lucid analysis of the problem(s) and solution(s).

But, what can be done to let clients upgrade orders of magnitude faster
than today?
Move transport in user space inside the app? Else?
Post by Jonathan Morton
I do see your arguments. Let it be known that I didn't initiate the
ack-filter in Cake, though it does seem to work quite well.
With respect to BBR, I don't think it depends strongly on the return rate
of acks in themselves, but rather on the rate of sequence number advance
that they indicate. For this purpose, having the receiver emit sparser but
still regularly spaced acks would be better than having some middlebox
delete some less-predictable subset of them. So I think BBR could be a
good testbed for AckCC implementation, especially as it is inherently paced
and thus doesn't suffer from burstiness as a conventional ack-clocked TCP
might.
The real trouble with AckCC is that it requires implementation on the
client as well as the server. That's most likely why Google hasn't tried
it yet; there are no receivers in the wild that would give them valid data
on its effectiveness. Adding support in Linux would help here, but aside
from Android devices, Linux is only a relatively small proportion of
Google's client traffic - and Android devices are slow to pick up new
kernel features if they can't immediately turn it into a consumer-friendly
bullet point.
Meanwhile we have highly asymmetric last-mile links (10:1 is typical, 50:1
is occasionally seen), where a large fraction of upload bandwidth is
occupied by acks in order to fully utilise the download bandwidth in TCP.
Any concurrent upload flows have to compete with that dense ack flow, which
in various schemes is unfair to either the upload or the download
throughput.
That is a problem as soon as you have multiple users on the same link, eg.
a family household at the weekend. Thinning out those acks in response to
uplink congestion is a solution. Maybe not the best possible solution, but
a deployable one that works.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Eric Dumazet
2017-11-30 10:24:40 UTC
Permalink
Raw Message
I agree that TCP itself should generate ACK smarter, on receivers that
are lacking GRO. (TCP sends at most one ACK per GRO packets, that is
why we did not feel an urgent need for better ACK generation)

It is actually difficult task, because it might need an additional
timer, and we were reluctant adding extra complexity for that.

An additional point where huge gains are possible is to add TSO
autodefer while in recovery. Lacking TSO auto defer explains why TCP
flows enter a degenerated behavior, re-sending 1-MSS packets in
response to SACK flood.
I do see your arguments.  Let it be known that I didn't initiate the
ack-filter in Cake, though it does seem to work quite well.
With respect to BBR, I don't think it depends strongly on the return
rate of acks in themselves, but rather on the rate of sequence number
advance that they indicate.  For this purpose, having the receiver
emit sparser but still regularly spaced acks would be better than
having some middlebox delete some less-predictable subset of them. 
So I think BBR could be a good testbed for AckCC implementation,
especially as it is inherently paced and thus doesn't suffer from
burstiness as a conventional ack-clocked TCP might.
The real trouble with AckCC is that it requires implementation on the
client as well as the server.  That's most likely why Google hasn't
tried it yet; there are no receivers in the wild that would give them
valid data on its effectiveness.  Adding support in Linux would help
here, but aside from Android devices, Linux is only a relatively
small proportion of Google's client traffic - and Android devices are
slow to pick up new kernel features if they can't immediately turn it
into a consumer-friendly bullet point.
Meanwhile we have highly asymmetric last-mile links (10:1 is typical,
50:1 is occasionally seen), where a large fraction of upload
bandwidth is occupied by acks in order to fully utilise the download
bandwidth in TCP.  Any concurrent upload flows have to compete with
that dense ack flow, which in various schemes is unfair to either the
upload or the download throughput.
That is a problem as soon as you have multiple users on the same
link, eg. a family household at the weekend.  Thinning out those acks
in response to uplink congestion is a solution.  Maybe not the best
possible solution, but a deployable one that works.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Mikael Abrahamsson
2017-11-30 13:04:45 UTC
Permalink
Raw Message
Post by Eric Dumazet
I agree that TCP itself should generate ACK smarter, on receivers that
are lacking GRO. (TCP sends at most one ACK per GRO packets, that is why
we did not feel an urgent need for better ACK generation)
Could you elaborate a bit more on the practical implications of the above
text? What is the typical GRO size used when doing gigabit ethernet
transmissions?

So if we're receiving 70kPPS of 1500 byte packets containing 1460 MSS
sized packet (~100 megabyte/s), what would a typical ACK rate be in that
case?

In response to some other postings here, my question regarding "is 35kPPS
really needed" my proposal is not "let's send 50 PPS of ACKs". My proposal
is if we can't come up with a smarter algorithm than something from the
90ties that says "let's send one ACK per 2*MSS" when we today have
magnitudes higher rates of forwarding. Also, on for instance DOCSIS
networks then you're going to get several ACKs back-to-back anyway
(because if they're not pruned by the DOCSIS network, they're anyway sent
in "bursts" within a single DOCSIS transmit opportunity), so imagining
that 35kPPS gives you higher resolution than 1kPPS of ACKs is just an
illusion.

So if GRO results in (I'm just speculating here) "we're only sending one
ACK per X kilobytes received if the packets arrived in the same
millisecond" and X is in the 16-64 kilobyte range, then that's fine by me.

Any network worth anything should be able to smooth out "bursts" of 16-64
kilobytes at line rate anyway, in case of egress and the line rate there
is lower than the sending end is transmitting packets at.
--
Mikael Abrahamsson email: ***@swm.pp.se
Eric Dumazet
2017-11-30 15:51:56 UTC
Permalink
Raw Message
Post by Mikael Abrahamsson
Post by Eric Dumazet
I agree that TCP itself should generate ACK smarter, on receivers
that 
are lacking GRO. (TCP sends at most one ACK per GRO packets, that
is why 
we did not feel an urgent need for better ACK generation)
Could you elaborate a bit more on the practical implications of the
above 
text? What is the typical GRO size used when doing gigabit ethernet 
transmissions?
Assuming NAPI handler receives a big packet train in one go [1], GRO
packets can be full size (45 MSS -> 65160 bytes of payload assuming
1448 bytes per frame)

[1] GRO engine has an opt-in high res timer helping to extend NAPI poll
if desired. https://git.kernel.org/pub/scm/linux/kernel/git/davem/net-n
ext.git/commit/?id=3b47d30396bae4f0bd1ff0dbcd7c4f5077e7df4e
Post by Mikael Abrahamsson
So if we're receiving 70kPPS of 1500 byte packets containing 1460
MSS 
sized packet (~100 megabyte/s), what would a typical ACK rate be in
that 
case?
1) Assuming receiver handles GRO.
2) Assuming few PSH flag set on incoming frames.
3) A default GRO engine on a 10Gbit NIC would probably not aggregate
packets, since 14 usec delay between each packet is too big to let NAPI
handler catch more than 1 packet per NIC RX interrupt.

But setting /sys/class/net/ethX/gro_flush_timeout to 14000 would allow
to build full size GRO packets (45 MSS)

-> TCP receiver would then send 1555 ACK per second instead of 70,000
Post by Mikael Abrahamsson
In response to some other postings here, my question regarding "is
35kPPS 
really needed" my proposal is not "let's send 50 PPS of ACKs". My
proposal 
is if we can't come up with a smarter algorithm than something from
the 
90ties that says "let's send one ACK per 2*MSS" when we today have 
magnitudes higher rates of forwarding. Also, on for instance DOCSIS 
networks then you're going to get several ACKs back-to-back anyway 
(because if they're not pruned by the DOCSIS network, they're anyway
sent 
in "bursts" within a single DOCSIS transmit opportunity), so
imagining 
that 35kPPS gives you higher resolution than 1kPPS of ACKs is just
an 
illusion.
So if GRO results in (I'm just speculating here) "we're only sending
one 
ACK per X kilobytes received if the packets arrived in the same 
millisecond" and X is in the 16-64 kilobyte range, then that's fine by me.
Any network worth anything should be able to smooth out "bursts" of
16-64 
kilobytes at line rate anyway, in case of egress and the line rate
there 
is lower than the sending end is transmitting packets at.
David Lang
2017-12-01 00:28:16 UTC
Permalink
Raw Message
35K PPS of acks is insane, one ack every ms is FAR more than enough to do 'fast
recovery', and outside the datacenter, one ack per 10ms is probably more than
enough.

Assuming something that's not too assymetric, thinning out the acks may not make
any difference in the transfer rate of a single data flow in one direction, but
if you step back and realize that there may be a need to transfer data in the
other direction, things change here.

If you have a fully symmetrical link, and are maxing it out in both direction,
going from 35K PPs of aks competing with data packets and gonig down to 1k PPS
or 100 PPS (or 10 PPS) would result in a noticable improvement in the flow that
the acks are competing against.

Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.

David Lang
Jan Ceuleers
2017-12-01 07:09:35 UTC
Permalink
Raw Message
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle
'upstream' paths.
Nobody has said it so I will: on wifi-connected endpoints the upstream
acks also compete for airtime with the downstream flow.
Toke Høiland-Jørgensen
2017-12-01 12:53:51 UTC
Permalink
Raw Message
Post by Jan Ceuleers
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.
Nobody has said it so I will: on wifi-connected endpoints the upstream
acks also compete for airtime with the downstream flow.
There's a related discussion going on over on the make-wifi-fast list
related to the FastACK scheme proposed by Meraki at this year's IMC:

https://conferences.sigcomm.org/imc/2017/papers/imc17-final203.pdf

It basically turns link-layer ACKs into upstream TCP ACKs (and handles
some of the corner cases resulting from this) and also seems to contain
an ACK compression component.

-Toke
Luca Muscariello
2017-12-01 13:17:07 UTC
Permalink
Raw Message
If I understand the text right, FastACK runs in the AP and generates an ACK
on behalf (or despite) of the TCP client end.
Then, it decimates dupACKs.

This means that there is a stateful connection tracker in the AP. Not so
simple.
It's almost, not entirely though, a TCP proxy doing Split TCP.
Post by Toke Høiland-Jørgensen
Post by Jan Ceuleers
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.
Nobody has said it so I will: on wifi-connected endpoints the upstream
acks also compete for airtime with the downstream flow.
There's a related discussion going on over on the make-wifi-fast list
https://conferences.sigcomm.org/imc/2017/papers/imc17-final203.pdf
It basically turns link-layer ACKs into upstream TCP ACKs (and handles
some of the corner cases resulting from this) and also seems to contain
an ACK compression component.
-Toke
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Toke Høiland-Jørgensen
2017-12-01 13:40:40 UTC
Permalink
Raw Message
Post by Luca Muscariello
If I understand the text right, FastACK runs in the AP and generates an ACK
on behalf (or despite) of the TCP client end.
Then, it decimates dupACKs.
This means that there is a stateful connection tracker in the AP. Not so
simple.
It's almost, not entirely though, a TCP proxy doing Split TCP.
Yeah, it's very much stateful, and tied closely to both TCP and the MAC
layer. So it has all the usual middlebox issues as far as that is
concerned... Also, APs need to transfer state between each other when
the client roams.

It does increase single-flow TCP throughput by up to a factor of two,
though... Which everyone knows is the most important benchmark number ;)

-Toke
Dave Taht
2017-12-01 17:42:13 UTC
Permalink
Raw Message
Post by Toke Høiland-Jørgensen
Post by Luca Muscariello
If I understand the text right, FastACK runs in the AP and generates an ACK
on behalf (or despite) of the TCP client end.
Then, it decimates dupACKs.
This means that there is a stateful connection tracker in the AP. Not so
simple.
It's almost, not entirely though, a TCP proxy doing Split TCP.
Yeah, it's very much stateful, and tied closely to both TCP and the MAC
layer. So it has all the usual middlebox issues as far as that is
concerned... Also, APs need to transfer state between each other when
the client roams.
It does increase single-flow TCP throughput by up to a factor of two,
though... Which everyone knows is the most important benchmark number ;)
Were you always as cynical as I am?

I'd like to compare (eventually) what we are trying with cake's new ack
filter here, which at least doesn't lie to the endpoint.

my guess, however, would be that the media access negotiation will
dominate the cost, and savings from (say) reducing 10 acks to 1 would
only be somewhere in the 5-20% range, for simple benchmarks.

I think we might get a better rrul result, however, as we'd be able to
pack more big flows into a given aggregate, with less acks there.
Post by Toke Høiland-Jørgensen
-Toke
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Juliusz Chroboczek
2017-12-01 20:39:29 UTC
Permalink
Raw Message
Post by Dave Taht
Post by Toke Høiland-Jørgensen
It does increase single-flow TCP throughput by up to a factor of two,
though... Which everyone knows is the most important benchmark number ;)
Were you always as cynical as I am?
(Giggle)

Dave, you've always underestimated Toke ;-)
Bob McMahon
2017-12-03 05:20:33 UTC
Permalink
Raw Message
I'm skeptical that this would improve single stream throughput by a factor
of two. The larger RTT would drive larger aggregations and it's
aggregation that scales peak average throughput.

Also, the time difference between the 802.11 ack and the client network
stack writing the TCP ack would probably be in the 100s of microseconds
(mileage will vary.) So it's the client's media access that will drive the
increase in RTT. It might be preferred to modify EDCA parameters to
reduce media access latencies for TCP acks rather than spoof them.

Bob
Post by Juliusz Chroboczek
Post by Dave Taht
Post by Toke Høiland-Jørgensen
It does increase single-flow TCP throughput by up to a factor of two,
though... Which everyone knows is the most important benchmark number ;)
Were you always as cynical as I am?
(Giggle)
Dave, you've always underestimated Toke ;-)
_______________________________________________
Make-wifi-fast mailing list
https://lists.bufferbloat.net/listinfo/make-wifi-fast
Juliusz Chroboczek
2017-12-03 10:35:12 UTC
Permalink
Raw Message
It might be preferred to modify EDCA parameters to reduce media access
latencies for TCP acks rather than spoof them.
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?

-- Juliusz
Jan Ceuleers
2017-12-03 11:40:04 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?
My understanding is that the issue that gave rise to this discussion was
concerned with upstream bandwidth conservation in the uplink of a DOCSIS
network by the cable modem dropping a large percentage of upstream TCP ACKs.

One element of that discussion was the question as to whether it was OK
for middleboxes (such as in this case cable modems) to reduce the number
of TCP ACKs, or whether instead the TCP stacks in the endpoints should
be made to send fewer such ACKs in the first place.

I then added more confusion by saying that in the case of wifi-connected
endpoints the upstream TCP ACKs also compete for airtime with the
downstream flow. Of course this no longer has anything to do with the
cable modem.
Juliusz Chroboczek
2017-12-03 13:57:17 UTC
Permalink
Raw Message
Post by Jan Ceuleers
Post by Juliusz Chroboczek
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?
My understanding is that the issue that gave rise to this discussion was
concerned with upstream bandwidth conservation in the uplink of a DOCSIS
network by the cable modem dropping a large percentage of upstream TCP ACKs.
Ok, that's what I thought. I'm glad we agree that WiFi is a different issue.

A TCP Ack is 40 bytes. A data packet is up to 1500 bytes.

As far as I know, DOCSIS has an asymmetry factor that is between 4 and 10,
depending on the deployment. With worst case asymmetry being 10, this
means that you can send an Ack for every data packet with 400 byte data
packets, every second data packet with 200 byte data packets. If the
asymmetry is a more reasonable 4, then the figures are 100 and 50
respectively.

Try as I might, I fail to see the problem. Are we advocating deploying
TCP-aware middleboxes, with all the problems that entails, in order to
work around a problem that doesn't exist?

-- Juliusz
Mikael Abrahamsson
2017-12-03 14:07:24 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
As far as I know, DOCSIS has an asymmetry factor that is between 4 and 10,
depending on the deployment. With worst case asymmetry being 10, this
I can buy 300/10 megabit/s access from my cable provider. So that's a lot
worse. My cable box has 16 downstream channels, and 4 upstream ones. Each
channel is TDM based, and there is some kind of scheduler granting sending
opportunities for each channel to each modem, as needed. I'm not a DOCSIS
expert.
Post by Juliusz Chroboczek
means that you can send an Ack for every data packet with 400 byte data
packets, every second data packet with 200 byte data packets. If the
asymmetry is a more reasonable 4, then the figures are 100 and 50
respectively.
Try as I might, I fail to see the problem. Are we advocating deploying
TCP-aware middleboxes, with all the problems that entails, in order to
work around a problem that doesn't exist?
If I understand correctly, DOCSIS has ~1ms sending opportunities upstream.
So sending more than 1kPPS of ACKs is meaningless, as these ACKs will just
come back to back at wire-speed as the CMTS receives them from the modem
in chunks. So instead, the cable modem just deletes all the sequential
ACKs and doesn't even send these back-to-back ones.

LTE works the same, it's also frequency divided and TDM, so I can see the
same benefit there of culling sequential ACKs sitting there in the buffer.
I don't know if this is done though.

I've seen people I think are involved in TCP design. They seem to be under
the impression that more ACKs give higher resolution and granularity to
TCP. My postulation is that this is commonly false because of how the
network access is designed and how also the NICs are designed (the
transmit/receive offloading). So sending 35kPPS of ACKs for a gigabit/s
transfer is just inefficient and shouldn't be done. I would prefer if end
points would send less ACKs instead of the network killing them.

And the network does kill them, as we have seen. Because any novice
network access technology designer can say "oh, having 16 sequential ACKs
here in my buffer, sitting waiting to get sent, is just useless
information. Let's kill the 15 first ones."
--
Mikael Abrahamsson email: ***@swm.pp.se
Juliusz Chroboczek
2017-12-03 19:53:32 UTC
Permalink
Raw Message
Post by Mikael Abrahamsson
I can buy 300/10 megabit/s access from my cable provider.
Don't!
Post by Mikael Abrahamsson
If I understand correctly, DOCSIS has ~1ms sending opportunities
upstream. So sending more than 1kPPS of ACKs is meaningless, as these ACKs
will just come back to back at wire-speed as the CMTS receives them from
the modem in chunks. So instead, the cable modem just deletes all the
sequential ACKs and doesn't even send these back-to-back ones.
If true -- then it's horrible.
Post by Mikael Abrahamsson
LTE works the same, it's also frequency divided and TDM, so I can see the
same benefit there of culling sequential ACKs sitting there in the
buffer. I don't know if this is done though.
I cannot find anything about Ack compression in LTE. (The PDCP protocol
does header compression, so that's the place I'm looking.)

-- Juliusz
Ryan Mounce
2017-12-03 14:09:17 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
Post by Jan Ceuleers
Post by Juliusz Chroboczek
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?
My understanding is that the issue that gave rise to this discussion was
concerned with upstream bandwidth conservation in the uplink of a DOCSIS
network by the cable modem dropping a large percentage of upstream TCP ACKs.
Ok, that's what I thought. I'm glad we agree that WiFi is a different issue.
A TCP Ack is 40 bytes. A data packet is up to 1500 bytes.
As far as I know, DOCSIS has an asymmetry factor that is between 4 and 10,
depending on the deployment. With worst case asymmetry being 10, this
means that you can send an Ack for every data packet with 400 byte data
packets, every second data packet with 200 byte data packets. If the
asymmetry is a more reasonable 4, then the figures are 100 and 50
respectively.
Many would kill for a 10:1 DOCSIS connection. 50:1 is not rare, and I
have personally been subscribed to a near 100:1 service.

Either way, the issue is not so much ACKs from downloads on an
otherwise idle link. The real issue is when the ACKs are contending
with a file upload, in this case download speeds will suffer if ACKs
are naively tail-dropped. Recovering extra bandwidth for the file
upload can be a happy side-effect.

You're also only counting IP packet length. The DOCSIS shaper deals
with ethernet frames so 58 / 1518 bytes.
Post by Juliusz Chroboczek
Try as I might, I fail to see the problem. Are we advocating deploying
TCP-aware middleboxes, with all the problems that entails, in order to
work around a problem that doesn't exist?
-- Juliusz
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Regards,
Ryan Mounce
Juliusz Chroboczek
2017-12-03 19:54:40 UTC
Permalink
Raw Message
Post by Ryan Mounce
Many would kill for a 10:1 DOCSIS connection. 50:1 is not rare, and I
have personally been subscribed to a near 100:1 service.
Some people should not be allowed to design networks.
Post by Ryan Mounce
The DOCSIS shaper deals with ethernet frames so 58 / 1518 bytes.
Could you please point me to details of the DOCSIS shaper?

-- Juliusz
Sebastian Moeller
2017-12-03 20:14:20 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
Post by Ryan Mounce
Many would kill for a 10:1 DOCSIS connection. 50:1 is not rare, and I
have personally been subscribed to a near 100:1 service.
Some people should not be allowed to design networks.
Post by Ryan Mounce
The DOCSIS shaper deals with ethernet frames so 58 / 1518 bytes.
Could you please point me to details of the DOCSIS shaper?
the relevant section from the Docsis standard (http://www.cablelabs.com/specification/docsis-3-0-mac-and-upper-layer-protocols-interface-specification/):

"C.2.2.7.2 Maximum Sustained Traffic Rate 632 This parameter is the rate parameter R of a token-bucket-based rate limit for packets. R is expressed in bits per second, and MUST take into account all MAC frame data PDU of the Service Flow from the byte following the MAC header HCS to the end of the CRC, including every PDU in the case of a Concatenated MAC Frame. This parameter is applied after Payload Header Suppression; it does not include the bytes suppressed for PHS. The number of bytes forwarded (in bytes) is limited during any time interval T by Max(T), as described in the expression: Max(T) = T * (R / 8) + B, (1) where the parameter B (in bytes) is the Maximum Traffic Burst Configuration Setting (refer to Annex C.2.2.7.3). NOTE: This parameter does not limit the instantaneous rate of the Service Flow. The specific algorithm for enforcing this parameter is not mandated here. Any implementation which satisfies the above equation is conformant. In particular, the granularity of enforcement and the minimum implemented value of this parameter are vendor specific. The CMTS SHOULD support a granularity of at most 100 kbps. The CM SHOULD support a granularity of at most 100 kbps. NOTE: If this parameter is omitted or set to zero, then there is no explicitly-enforced traffic rate maximum. This field specifies only a bound, not a guarantee that this rate is available."

So in essence DOCSIS users need to only account for 18 Bytes of ethernet overhead in both ingress and egress directions under non-congested conditions.
Post by Juliusz Chroboczek
-- Juliusz
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
Dave Taht
2017-12-03 22:27:53 UTC
Permalink
Raw Message
Post by Sebastian Moeller
Post by Juliusz Chroboczek
Post by Ryan Mounce
Many would kill for a 10:1 DOCSIS connection. 50:1 is not rare, and I
have personally been subscribed to a near 100:1 service.
Some people should not be allowed to design networks.
The upstream/downstream problem over long distances has been
problematic for dsl (phone line) and
cable (coax) deployments. The head-ends have much greater control over
the signal strengths than the
(usually much cheaper)

Gpon fiber is also commonly sold in 1Gbit/100Mbit modes. Testing on a
GPON network showed about
80ms worth of buffering in the ONT - which we can get rid of entirely, in cake.
Post by Sebastian Moeller
Post by Juliusz Chroboczek
Post by Ryan Mounce
The DOCSIS shaper deals with ethernet frames so 58 / 1518 bytes.
Could you please point me to details of the DOCSIS shaper?
"C.2.2.7.2 Maximum Sustained Traffic Rate 632 This parameter is the rate parameter R of a token-bucket-based rate limit for packets. R is expressed in bits per second, and MUST take into account all MAC frame data PDU of the Service Flow from the byte following the MAC header HCS to the end of the CRC, including every PDU in the case of a Concatenated MAC Frame. This parameter is applied after Payload Header Suppression; it does not include the bytes suppressed for PHS. The number of bytes forwarded (in bytes) is limited during any time interval T by Max(T), as described in the expression: Max(T) = T * (R / 8) + B, (1) where the parameter B (in bytes) is the Maximum Traffic Burst Configuration Setting (refer to Annex C.2.2.7.3). NOTE: This parameter does not limit the instantaneous rate of the Service Flow. The specific algorithm for enforcing this parameter is not mandated here. Any implementation which satisfies the above equation is conformant. In particular, the granularity of enforcement and the minimum implemented value of this parameter are vendor specific. The CMTS SHOULD support a granularity of at most 100 kbps. The CM SHOULD support a granularity of at most 100 kbps. NOTE: If this parameter is omitted or set to zero, then there is no explicitly-enforced traffic rate maximum. This field specifies only a bound, not a guarantee that this rate is available."
So in essence DOCSIS users need to only account for 18 Bytes of ethernet overhead in both ingress and egress directions under non-congested conditions.
Also, cake, as a deficit mode shaper, it is the opposite of how htb
functions in terms of bursts. TB tries to make up
for bandwidth you should have, verses cake which gives you the
bandwidth you have "right now".

This lets us set the shaper much closer (seemingly exact in the case
of docsis, atleast) to the actual configured TB rate (with better
fq/aqm queue management)

I just submitted an initial patch for cake to net-next after a huge
round of testing.
Post by Sebastian Moeller
Post by Juliusz Chroboczek
-- Juliusz
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Sent from my Android device with K-9 Mail. Please excuse my brevity.
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Robert Bradley
2017-12-03 15:25:44 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
A TCP Ack is 40 bytes. A data packet is up to 1500 bytes.
As far as I know, DOCSIS has an asymmetry factor that is between 4 and 10,
depending on the deployment. With worst case asymmetry being 10, this
means that you can send an Ack for every data packet with 400 byte data
packets, every second data packet with 200 byte data packets. If the
asymmetry is a more reasonable 4, then the figures are 100 and 50
respectively.
I currently have 230 Mb/s down to 12.7 Mb/s up, so about an 18:1 ratio.
That's roughly an ACK for every 750 byte packet.
--
Robert Bradley
Dave Taht
2017-12-04 03:44:27 UTC
Permalink
Raw Message
Post by Jan Ceuleers
Post by Juliusz Chroboczek
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?
My understanding is that the issue that gave rise to this discussion was
concerned with upstream bandwidth conservation in the uplink of a DOCSIS
network by the cable modem dropping a large percentage of upstream TCP ACKs.
One element of that discussion was the question as to whether it was OK
for middleboxes (such as in this case cable modems) to reduce the number
of TCP ACKs, or whether instead the TCP stacks in the endpoints should
be made to send fewer such ACKs in the first place.
I then added more confusion by saying that in the case of wifi-connected
endpoints the upstream TCP ACKs also compete for airtime with the
downstream flow. Of course this no longer has anything to do with the
cable modem.
More generally, the case where you have a queue containing acks, stored
up for whatever reason (congestion, media access, asymmetry), is a
chance for a middlebox or host to do something "smarter" to thin them
out.

Acks don't respond to conventional congestion control mechanisms anyway.

There is another case (that I don't support) where you would try to
filter out acks on the fly without a queue (similar to how a policer
works). The flaws of this approach are many, including tail loss,
which the concept of filtering down (reducing?) a queue, doesn't have.

fq_codel has a tendency to gather up flows into a quantum (usually 1514
bytes), which translates out to 22 ipv4 acks before it will switch flows.

The cake implementation will always deliver the lastmost ack packet, and
also has some compensations for stuff in slow start.

(it could use a more formal state machine, and perhaps tuning out the
sparse flow optimization, and more testing. It certainly is not fast
code, but still cheaper than the hashing bits in cake)
Post by Jan Ceuleers
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
David Collier-Brown
2017-12-04 14:38:31 UTC
Permalink
Raw Message
Post by Dave Taht
More generally, the case where you have a queue containing acks, stored
up for whatever reason (congestion, media access, asymmetry), is a
chance for a middlebox or host to do something "smarter" to thin them
out.
Acks don't respond to conventional congestion control mechanisms anyway.
There is another case (that I don't support) where you would try to
filter out acks on the fly without a queue (similar to how a policer
works). The flaws of this approach are many, including tail loss,
which the concept of filtering down (reducing?) a queue, doesn't have.
Taking a very high-level view of this discussion, the times you want to
change a protocol or add a 'network optimizer" are when enough time has
passed that the original requirements don't describe what you want any more.

In a previous life I did some work on the optimization (by remote
proxying) of the SMB protocol used by Samba. It was very desirable, but
at the cost of continuing to support a protocol that did the wrong
thing, and kludging it with additional middleware.  In effect, making
your new system dependent on a bug in the old one.

Eventually we said the heck with it, and sat Samba on top of a different
protocol entirely, one which worked well over non-local links. That
concentrate the impedance matching in Samba, not in code I had to
maintain in synchronization with a bug (;-))

--dave
--
David Collier-Brown, | Always do right. This will gratify
System Programmer and Author | some people and astonish the rest
***@spamcop.net | -- Mark Twain
Juliusz Chroboczek
2017-12-04 15:44:38 UTC
Permalink
Raw Message
Post by David Collier-Brown
In a previous life I did some work on the optimization (by remote
proxying) of the SMB protocol used by Samba [...] Eventually we said
the heck with it, and sat Samba on top of a different protocol entirely,
The audience are waiting with held breath for more details.

-- Juliusz
David Collier-Brown
2017-12-04 17:17:48 UTC
Permalink
Raw Message
Post by Juliusz Chroboczek
Post by David Collier-Brown
In a previous life I did some work on the optimization (by remote
proxying) of the SMB protocol used by Samba [...] Eventually we said
the heck with it, and sat Samba on top of a different protocol entirely,
The audience are waiting with held breath for more details.
-- Juliusz
They aren't discussable in polite company. Way too much cursing (;-))

Joking aside, that was definitely a case where we said "don't go
there".   To the best of my knowledge, there are two network
optimization products that do SMB, so it's physically possible.  In our
opinion, it was better to use the SMB protocol locally and a different,
cached,  protocol over a wide-area network.  I actually prototyped it
with Solaris NFS and cachefs, and was pleasantly surprised it worked for
a single-writer case.

--dave
--
David Collier-Brown, | Always do right. This will gratify
System Programmer and Author | some people and astonish the rest
***@spamcop.net | -- Mark Twain
Bob McMahon
2017-12-03 19:04:45 UTC
Permalink
Raw Message
My understanding per the thread is a last hop wifi link. I could be wrong
though.

Bob
Post by Juliusz Chroboczek
It might be preferred to modify EDCA parameters to reduce media access
latencies for TCP acks rather than spoof them.
I'm lost here. What exact problem is the ACK hack supposed to work
around? Ridiculous amount of asymmetry in the last-hop WiFi link, or
outrageous amounts of asymmetry in a transit link beyond the last hop?
-- Juliusz
Bob McMahon
2017-12-01 21:17:59 UTC
Permalink
Raw Message
802.11 acks are packet or ampdu driven while tcp, being a byte protocol,
acks bytes. Aligning these may not be straightforward. We test with
different read() rates on the wifi clients as TCP is supposed to flow
control the source's writes() as well. Wifi clients are starting to align
their sleep cycles with "natural" periodicity in traffic so having larger
aggregates can help both peak average throughput as well as power
consumption. It's not obvious with Wifi that a faster RTT is always
better. (Reminds me of the early days of NASA where many designed to
reduce weight without keeping in account structural integrity, shave a few
grams and lose a rocket.)

Bob
Post by Dave Taht
Post by Toke Høiland-Jørgensen
Post by Luca Muscariello
If I understand the text right, FastACK runs in the AP and generates an
ACK
Post by Toke Høiland-Jørgensen
Post by Luca Muscariello
on behalf (or despite) of the TCP client end.
Then, it decimates dupACKs.
This means that there is a stateful connection tracker in the AP. Not so
simple.
It's almost, not entirely though, a TCP proxy doing Split TCP.
Yeah, it's very much stateful, and tied closely to both TCP and the MAC
layer. So it has all the usual middlebox issues as far as that is
concerned... Also, APs need to transfer state between each other when
the client roams.
It does increase single-flow TCP throughput by up to a factor of two,
though... Which everyone knows is the most important benchmark number ;)
Were you always as cynical as I am?
I'd like to compare (eventually) what we are trying with cake's new ack
filter here, which at least doesn't lie to the endpoint.
my guess, however, would be that the media access negotiation will
dominate the cost, and savings from (say) reducing 10 acks to 1 would
only be somewhere in the 5-20% range, for simple benchmarks.
I think we might get a better rrul result, however, as we'd be able to
pack more big flows into a given aggregate, with less acks there.
Post by Toke Høiland-Jørgensen
-Toke
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Make-wifi-fast mailing list
https://lists.bufferbloat.net/listinfo/make-wifi-fast
Luca Muscariello
2017-12-01 13:22:49 UTC
Permalink
Raw Message
I think only IPSEC would be a problem for fastACK but not TLS.
"FastACK also relies on packet inspection, and will not work when
payload is encrypted. However, in our networks, we do not currently
see an extensive use of encryption techniques like IPSec."
But what about TLS ?
As for me, this technology will never work in most cases.
Best regards,
Lukonin Kirill.
Post by Toke Høiland-Jørgensen
Post by Jan Ceuleers
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.
Nobody has said it so I will: on wifi-connected endpoints the upstream
acks also compete for airtime with the downstream flow.
There's a related discussion going on over on the make-wifi-fast list
https://conferences.sigcomm.org/imc/2017/papers/imc17-final203.pdf
It basically turns link-layer ACKs into upstream TCP ACKs (and handles
some of the corner cases resulting from this) and also seems to contain
an ACK compression component.
-Toke
_______________________________________________
Make-wifi-fast mailing list
https://lists.bufferbloat.net/listinfo/make-wifi-fast
--
Best Regards,
Lukonin Kirill
_______________________________________________
Make-wifi-fast mailing list
https://lists.bufferbloat.net/listinfo/make-wifi-fast
Simon Barber
2017-12-11 17:42:05 UTC
Permalink
Raw Message
TLS works over TCP, so the TCP headers are not encrypted.

Simon

Sent with AquaMail for Android
http://www.aqua-mail.com
"FastACK also relies on packet inspection, and will not work when
payload is encrypted. However, in our networks, we do not currently
see an extensive use of encryption techniques like IPSec."
But what about TLS ?
As for me, this technology will never work in most cases.
Best regards,
Lukonin Kirill.
Post by Toke Høiland-Jørgensen
Post by Jan Ceuleers
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.
Nobody has said it so I will: on wifi-connected endpoints the upstream
acks also compete for airtime with the downstream flow.
There's a related discussion going on over on the make-wifi-fast list
https://conferences.sigcomm.org/imc/2017/papers/imc17-final203.pdf
It basically turns link-layer ACKs into upstream TCP ACKs (and handles
some of the corner cases resulting from this) and also seems to contain
an ACK compression component.
-Toke
_______________________________________________
Make-wifi-fast mailing list
https://lists.bufferbloat.net/listinfo/make-wifi-fast
--
Best Regards,
Lukonin Kirill
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Sebastian Moeller
2017-12-01 08:45:42 UTC
Permalink
Raw Message
Hi All,

you do realize that the worst case is going to stay at 35KPPS? If we assume simply that the 100Mbps download rate is not created by a single flow but by many flows (say 70K flows) the discussed ACK frequency reduction schemes will not work that well. So ACK thinning is a nice optimization, but will not help the fact that some ISPs/link technologies simply are asymmetric and the user will suffer under some traffic conditions. Now the 70K flow example is too extreme, but the fact is at hight flow number with sparse flows (so fewer ACKs per flow in the queue and fewer ACKs per flow reaching the end NIC in a GRO-collection interval (I naively assume there is a somewhat fixed but small interval in which packets of the same flow are collected for GRO)) there will be problems. (Again, I am all for allowing the end user to configure ACK filtering thinning, but I would rather see ISPs sell less imbalanced links ;) )

Best Regards
Sebastian
35K PPS of acks is insane, one ack every ms is FAR more than enough to do 'fast recovery', and outside the datacenter, one ack per 10ms is probably more than enough.
Assuming something that's not too assymetric, thinning out the acks may not make any difference in the transfer rate of a single data flow in one direction, but if you step back and realize that there may be a need to transfer data in the other direction, things change here.
If you have a fully symmetrical link, and are maxing it out in both direction, going from 35K PPs of aks competing with data packets and gonig down to 1k PPS or 100 PPS (or 10 PPS) would result in a noticable improvement in the flow that the acks are competing against.
Stop thinking in terms of single-flow benchmarks and near idle 'upstream' paths.
David Lang
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Luca Muscariello
2017-12-01 10:45:19 UTC
Permalink
Raw Message
For highly asymmetric links, but also shared media like wifi, QUIC might be
a better playground for optimisations.
Not pervasive as TCP though and maybe off topic in this thread.

If the downlink is what one want to optimise, using FEC in the downstream,
in conjunction with flow control could be very effective.
No need to send ACK frequently and having something like FQ_codel in the
downstream would avoid fairness problems that might
happen though. I don't know if FEC is still in QUIC and used.


BTW, for wifi, the ACK stream can be compressed in aggregate of frames and
sent in bursts. This is similar to DOCSIS upstream.
I wonder if this is a phenomenon that is visible in recent WiFi or just
negligible.
Post by Sebastian Moeller
Hi All,
you do realize that the worst case is going to stay at 35KPPS? If we
assume simply that the 100Mbps download rate is not created by a single
flow but by many flows (say 70K flows) the discussed ACK frequency
reduction schemes will not work that well. So ACK thinning is a nice
optimization, but will not help the fact that some ISPs/link technologies
simply are asymmetric and the user will suffer under some traffic
conditions. Now the 70K flow example is too extreme, but the fact is at
hight flow number with sparse flows (so fewer ACKs per flow in the queue
and fewer ACKs per flow reaching the end NIC in a GRO-collection interval
(I naively assume there is a somewhat fixed but small interval in which
packets of the same flow are collected for GRO)) there will be problems.
(Again, I am all for allowing the end user to configure ACK filtering
thinning, but I would rather see ISPs sell less imbalanced links ;) )
Best Regards
Sebastian
Post by David Lang
35K PPS of acks is insane, one ack every ms is FAR more than enough to
do 'fast recovery', and outside the datacenter, one ack per 10ms is
probably more than enough.
Post by David Lang
Assuming something that's not too assymetric, thinning out the acks may
not make any difference in the transfer rate of a single data flow in one
direction, but if you step back and realize that there may be a need to
transfer data in the other direction, things change here.
Post by David Lang
If you have a fully symmetrical link, and are maxing it out in both
direction, going from 35K PPs of aks competing with data packets and gonig
down to 1k PPS or 100 PPS (or 10 PPS) would result in a noticable
improvement in the flow that the acks are competing against.
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle
'upstream' paths.
Post by David Lang
David Lang
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Dave Taht
2017-12-01 18:43:22 UTC
Permalink
Raw Message
For highly asymmetric links, but also shared media like wifi, QUIC might be a
better playground for optimisations.
Not pervasive as TCP though and maybe off topic in this thread.
I happen to really like QUIC, but a netperf-style tool did not exist for
it when I last looked, last year.

Also getting to emulating DASH traffic is on my list.
If the downlink is what one want to optimise, using FEC in the downstream, in
conjunction with flow control could be very effective.
No need to send ACK frequently and having something like FQ_codel in the
downstream would avoid fairness problems that might
happen though. I don't know if FEC is still in QUIC and used.
BTW, for wifi, the ACK stream can be compressed in aggregate of frames and sent
in bursts. This is similar to DOCSIS upstream.
I wonder if this is a phenomenon that is visible in recent WiFi or just
negligible.
My guess is meraki deployed something and I think they are in in the top
5 in the enterprise market.

I see ubnt added airtime fairness (of some sort), recently.
Hi All,
you do realize that the worst case is going to stay at 35KPPS? If we assume
simply that the 100Mbps download rate is not created by a single flow but by
many flows (say 70K flows) the discussed ACK frequency reduction schemes
will not work that well. So ACK thinning is a nice optimization, but will
not help the fact that some ISPs/link technologies simply are asymmetric and
the user will suffer under some traffic conditions. Now the 70K flow example
is too extreme, but the fact is at hight flow number with sparse flows (so
fewer ACKs per flow in the queue and fewer ACKs per flow reaching the end
NIC in a GRO-collection interval (I naively assume there is a somewhat fixed
but small interval in which packets of the same flow are collected for GRO))
there will be problems. (Again, I am all for allowing the end user to
configure ACK filtering thinning, but I would rather see ISPs sell less
imbalanced links ;) )
Best Regards
Sebastian
Post by David Lang
35K PPS of acks is insane, one ack every ms is FAR more than enough to do
'fast recovery', and outside the datacenter, one ack per 10ms is probably
more than enough.
Post by David Lang
Assuming something that's not too assymetric, thinning out the acks may
not make any difference in the transfer rate of a single data flow in one
direction, but if you step back and realize that there may be a need to
transfer data in the other direction, things change here.
Post by David Lang
If you have a fully symmetrical link, and are maxing it out in both
direction, going from 35K PPs of aks competing with data packets and gonig
down to 1k PPS or 100 PPS (or 10 PPS) would result in a noticable
improvement in the flow that the acks are competing against.
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle 'upstream'
paths.
Post by David Lang
David Lang
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Luca Muscariello
2017-12-01 18:57:05 UTC
Permalink
Raw Message
https://www.cisco.com/c/en/us/products/collateral/wireless/aironet-3700-series/white-paper-c11-735947.html
Post by Dave Taht
Post by Luca Muscariello
For highly asymmetric links, but also shared media like wifi, QUIC might
be a
Post by Luca Muscariello
better playground for optimisations.
Not pervasive as TCP though and maybe off topic in this thread.
I happen to really like QUIC, but a netperf-style tool did not exist for
it when I last looked, last year.
Also getting to emulating DASH traffic is on my list.
Post by Luca Muscariello
If the downlink is what one want to optimise, using FEC in the
downstream, in
Post by Luca Muscariello
conjunction with flow control could be very effective.
No need to send ACK frequently and having something like FQ_codel in the
downstream would avoid fairness problems that might
happen though. I don't know if FEC is still in QUIC and used.
BTW, for wifi, the ACK stream can be compressed in aggregate of frames
and sent
Post by Luca Muscariello
in bursts. This is similar to DOCSIS upstream.
I wonder if this is a phenomenon that is visible in recent WiFi or just
negligible.
My guess is meraki deployed something and I think they are in in the top
5 in the enterprise market.
I see ubnt added airtime fairness (of some sort), recently.
Post by Luca Muscariello
Hi All,
you do realize that the worst case is going to stay at 35KPPS? If we
assume
Post by Luca Muscariello
simply that the 100Mbps download rate is not created by a single
flow but by
Post by Luca Muscariello
many flows (say 70K flows) the discussed ACK frequency reduction
schemes
Post by Luca Muscariello
will not work that well. So ACK thinning is a nice optimization, but
will
Post by Luca Muscariello
not help the fact that some ISPs/link technologies simply are
asymmetric and
Post by Luca Muscariello
the user will suffer under some traffic conditions. Now the 70K flow
example
Post by Luca Muscariello
is too extreme, but the fact is at hight flow number with sparse
flows (so
Post by Luca Muscariello
fewer ACKs per flow in the queue and fewer ACKs per flow reaching
the end
Post by Luca Muscariello
NIC in a GRO-collection interval (I naively assume there is a
somewhat fixed
Post by Luca Muscariello
but small interval in which packets of the same flow are collected
for GRO))
Post by Luca Muscariello
there will be problems. (Again, I am all for allowing the end user to
configure ACK filtering thinning, but I would rather see ISPs sell
less
Post by Luca Muscariello
imbalanced links ;) )
Best Regards
Sebastian
Post by David Lang
35K PPS of acks is insane, one ack every ms is FAR more than
enough to do
Post by Luca Muscariello
'fast recovery', and outside the datacenter, one ack per 10ms is
probably
Post by Luca Muscariello
more than enough.
Post by David Lang
Assuming something that's not too assymetric, thinning out the
acks may
Post by Luca Muscariello
not make any difference in the transfer rate of a single data flow
in one
Post by Luca Muscariello
direction, but if you step back and realize that there may be a need
to
Post by Luca Muscariello
transfer data in the other direction, things change here.
Post by David Lang
If you have a fully symmetrical link, and are maxing it out in both
direction, going from 35K PPs of aks competing with data packets and
gonig
Post by Luca Muscariello
down to 1k PPS or 100 PPS (or 10 PPS) would result in a noticable
improvement in the flow that the acks are competing against.
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle
'upstream'
Post by Luca Muscariello
paths.
Post by David Lang
David Lang
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Dave Taht
2017-12-01 19:36:12 UTC
Permalink
Raw Message
On Fri, Dec 1, 2017 at 10:57 AM, Luca Muscariello
Post by Luca Muscariello
https://www.cisco.com/c/en/us/products/collateral/wireless/aironet-3700-series/white-paper-c11-735947.html
Good news all over. I wonder what happens on cisco against the suite
of tests toke made available here:

https://www.cs.kau.se/tohojo/airtime-fairness/

People are getting some good results with this stuff:
https://forum.lede-project.org/t/ubiquiti-unifi-ac-mesh/4499/4
(however, I currently have 6 bricked ones that I need to recover, and
am having way more fun in simulation that I imagined I could ever
have)....
Post by Luca Muscariello
Post by Dave Taht
For highly asymmetric links, but also shared media like wifi, QUIC might be a
better playground for optimisations.
Not pervasive as TCP though and maybe off topic in this thread.
I happen to really like QUIC, but a netperf-style tool did not exist for
it when I last looked, last year.
Also getting to emulating DASH traffic is on my list.
If the downlink is what one want to optimise, using FEC in the downstream, in
conjunction with flow control could be very effective.
No need to send ACK frequently and having something like FQ_codel in the
downstream would avoid fairness problems that might
happen though. I don't know if FEC is still in QUIC and used.
BTW, for wifi, the ACK stream can be compressed in aggregate of frames and sent
in bursts. This is similar to DOCSIS upstream.
I wonder if this is a phenomenon that is visible in recent WiFi or just
negligible.
My guess is meraki deployed something and I think they are in in the top
5 in the enterprise market.
I see ubnt added airtime fairness (of some sort), recently.
Hi All,
you do realize that the worst case is going to stay at 35KPPS? If we assume
simply that the 100Mbps download rate is not created by a single flow but by
many flows (say 70K flows) the discussed ACK frequency reduction schemes
will not work that well. So ACK thinning is a nice optimization, but will
not help the fact that some ISPs/link technologies simply are asymmetric and
the user will suffer under some traffic conditions. Now the 70K flow example
is too extreme, but the fact is at hight flow number with sparse flows (so
fewer ACKs per flow in the queue and fewer ACKs per flow reaching the end
NIC in a GRO-collection interval (I naively assume there is a somewhat fixed
but small interval in which packets of the same flow are collected for GRO))
there will be problems. (Again, I am all for allowing the end user to
configure ACK filtering thinning, but I would rather see ISPs sell less
imbalanced links ;) )
Best Regards
Sebastian
Post by David Lang
35K PPS of acks is insane, one ack every ms is FAR more than
enough to do
'fast recovery', and outside the datacenter, one ack per 10ms is probably
more than enough.
Post by David Lang
Assuming something that's not too assymetric, thinning out the
acks may
not make any difference in the transfer rate of a single data flow in one
direction, but if you step back and realize that there may be a need to
transfer data in the other direction, things change here.
Post by David Lang
If you have a fully symmetrical link, and are maxing it out in
both
direction, going from 35K PPs of aks competing with data packets and gonig
down to 1k PPS or 100 PPS (or 10 PPS) would result in a noticable
improvement in the flow that the acks are competing against.
Post by David Lang
Stop thinking in terms of single-flow benchmarks and near idle
'upstream'
paths.
Post by David Lang
David Lang
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Neal Cardwell
2017-11-30 14:51:56 UTC
Permalink
Raw Message
Post by Eric Dumazet
I agree that TCP itself should generate ACK smarter, on receivers that
are lacking GRO. (TCP sends at most one ACK per GRO packets, that is
why we did not feel an urgent need for better ACK generation)
It is actually difficult task, because it might need an additional
timer, and we were reluctant adding extra complexity for that.
How about just using the existing delayed ACK timer, and just making the
delayed ACK logic a bit smarter? We could try using the existing logic and
timers, but using something adaptive instead of the magic "2" MSS received
to force an ACK.
Post by Eric Dumazet
An additional point where huge gains are possible is to add TSO
autodefer while in recovery. Lacking TSO auto defer explains why TCP
flows enter a degenerated behavior, re-sending 1-MSS packets in
response to SACK flood.
Yes, agreed. I suspect there is some simple heuristic that could be
implemented to allow TSO deferral for most packets sent in recovery. For
example, allowing TSO deferral once the number of packet bursts (TSO skbs)
sent in recovery is greater than some threshold. Perhaps TSO deferral would
be fine in Recovery if we have sent, say, 10 skbs, because at that point if
the ACK stream from the original flight dries up due to massive/tail loss,
we have probably sent enough data in the new flight in Recovery to ensure
some kind of ACKs come back to keep the ACK clock going.

neal
Post by Eric Dumazet
Post by Jonathan Morton
I do see your arguments. Let it be known that I didn't initiate the
ack-filter in Cake, though it does seem to work quite well.
With respect to BBR, I don't think it depends strongly on the return
rate of acks in themselves, but rather on the rate of sequence number
advance that they indicate. For this purpose, having the receiver
emit sparser but still regularly spaced acks would be better than
having some middlebox delete some less-predictable subset of them.
So I think BBR could be a good testbed for AckCC implementation,
especially as it is inherently paced and thus doesn't suffer from
burstiness as a conventional ack-clocked TCP might.
The real trouble with AckCC is that it requires implementation on the
client as well as the server. That's most likely why Google hasn't
tried it yet; there are no receivers in the wild that would give them
valid data on its effectiveness. Adding support in Linux would help
here, but aside from Android devices, Linux is only a relatively
small proportion of Google's client traffic - and Android devices are
slow to pick up new kernel features if they can't immediately turn it
into a consumer-friendly bullet point.
Meanwhile we have highly asymmetric last-mile links (10:1 is typical,
50:1 is occasionally seen), where a large fraction of upload
bandwidth is occupied by acks in order to fully utilise the download
bandwidth in TCP. Any concurrent upload flows have to compete with
that dense ack flow, which in various schemes is unfair to either the
upload or the download throughput.
That is a problem as soon as you have multiple users on the same
link, eg. a family household at the weekend. Thinning out those acks
in response to uplink congestion is a solution. Maybe not the best
possible solution, but a deployable one that works.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Eric Dumazet
2017-11-30 15:55:19 UTC
Permalink
Raw Message
Post by Neal Cardwell
Post by Eric Dumazet
I agree that TCP itself should generate ACK smarter, on receivers that
are lacking GRO. (TCP sends at most one ACK per GRO packets, that is
why we did not feel an urgent need for better ACK generation)
It is actually difficult task, because it might need an additional
timer, and we were reluctant adding extra complexity for that.
How about just using the existing delayed ACK timer, and just making
the delayed ACK logic a bit smarter? We could try using the existing
logic and timers, but using something adaptive instead of the magic
"2" MSS received to force an ACK.
Keep in mind some distros have HZ=250 or even HZ=100

So even a 'one jiffie' timer could add 10ms delay.

That is why I believe only a hrtimer could be used (and that would
imply CONFIG_HIGH_RES_TIMERS=y )

I am waiting for Anna-Maria Gleixner work (  hrtimer: Provide softirq
context hrtimers ) so that we can avoid a trip through a tasklet.
Post by Neal Cardwell
 
Post by Eric Dumazet
An additional point where huge gains are possible is to add TSO
autodefer while in recovery. Lacking TSO auto defer explains why TCP
flows enter a degenerated behavior, re-sending 1-MSS packets in
response to SACK flood.
Yes, agreed. I suspect there is some simple heuristic that could be
implemented to allow TSO deferral for most packets sent in recovery.
For example, allowing TSO deferral once the number of packet bursts
(TSO skbs) sent in recovery is greater than some threshold. Perhaps
TSO deferral would be fine in Recovery if we have sent, say, 10 skbs,
because at that point if the ACK stream from the original flight
dries up due to massive/tail loss, we have probably sent enough data
in the new flight in Recovery to ensure some kind of ACKs come back
to keep the ACK clock going.
neal
 
Post by Eric Dumazet
I do see your arguments.  Let it be known that I didn't initiate
the
ack-filter in Cake, though it does seem to work quite well.
With respect to BBR, I don't think it depends strongly on the
return
rate of acks in themselves, but rather on the rate of sequence
number
advance that they indicate.  For this purpose, having the
receiver
emit sparser but still regularly spaced acks would be better than
having some middlebox delete some less-predictable subset of
them. 
So I think BBR could be a good testbed for AckCC implementation,
especially as it is inherently paced and thus doesn't suffer from
burstiness as a conventional ack-clocked TCP might.
The real trouble with AckCC is that it requires implementation on
the
client as well as the server.  That's most likely why Google
hasn't
tried it yet; there are no receivers in the wild that would give
them
valid data on its effectiveness.  Adding support in Linux would
help
here, but aside from Android devices, Linux is only a relatively
small proportion of Google's client traffic - and Android devices
are
slow to pick up new kernel features if they can't immediately
turn it
into a consumer-friendly bullet point.
Meanwhile we have highly asymmetric last-mile links (10:1 is
typical,
50:1 is occasionally seen), where a large fraction of upload
bandwidth is occupied by acks in order to fully utilise the
download
bandwidth in TCP.  Any concurrent upload flows have to compete
with
that dense ack flow, which in various schemes is unfair to either
the
upload or the download throughput.
That is a problem as soon as you have multiple users on the same
link, eg. a family household at the weekend.  Thinning out those
acks
in response to uplink congestion is a solution.  Maybe not the
best
possible solution, but a deployable one that works.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Neal Cardwell
2017-11-30 15:57:07 UTC
Permalink
Raw Message
Post by Eric Dumazet
Post by Neal Cardwell
Post by Eric Dumazet
I agree that TCP itself should generate ACK smarter, on receivers that
are lacking GRO. (TCP sends at most one ACK per GRO packets, that is
why we did not feel an urgent need for better ACK generation)
It is actually difficult task, because it might need an additional
timer, and we were reluctant adding extra complexity for that.
How about just using the existing delayed ACK timer, and just making
the delayed ACK logic a bit smarter? We could try using the existing
logic and timers, but using something adaptive instead of the magic
"2" MSS received to force an ACK.
Keep in mind some distros have HZ=250 or even HZ=100
So even a 'one jiffie' timer could add 10ms delay.
Right, good point. I forgot about those cases. :-)

neal
Sebastian Moeller
2017-11-29 16:50:10 UTC
Permalink
Raw Message
Hi Mikael,
Post by Mikael Abrahamsson
Post by Sebastian Moeller
Well, ACK filtering/thinning is a simple trade-off: redundancy versus bandwidth. Since the RFCs say a receiver should acknoledge every second full MSS I think the decision whether to filter or not should be kept to
Why does it say to do this?
According to RFC 2525:
"2.13.

Name of Problem
Stretch ACK violation




Paxson, et. al. Informational [Page 40]

RFC 2525 TCP Implementation Problems March 1999



Classification
Congestion Control/Performance

Description
To improve efficiency (both computer and network) a data receiver
may refrain from sending an ACK for each incoming segment,
according to [
RFC1122
]. However, an ACK should not be delayed an
inordinate amount of time. Specifically, ACKs SHOULD be sent for
every second full-sized segment that arrives. If a second full-
sized segment does not arrive within a given timeout (of no more
than 0.5 seconds), an ACK should be transmitted, according to
[
RFC1122
]. A TCP receiver which does not generate an ACK for
every second full-sized segment exhibits a "Stretch ACK
Violation".

Significance
TCP receivers exhibiting this behavior will cause TCP senders to
generate burstier traffic, which can degrade performance in
congested environments. In addition, generating fewer ACKs
increases the amount of time needed by the slow start algorithm to
open the congestion window to an appropriate point, which
diminishes performance in environments with large bandwidth-delay
products. Finally, generating fewer ACKs may cause needless
retransmission timeouts in lossy environments, as it increases the
possibility that an entire window of ACKs is lost, forcing a
retransmission timeout.

Implications
When not in loss recovery, every ACK received by a TCP sender
triggers the transmission of new data segments. The burst size is
determined by the number of previously unacknowledged segments
each ACK covers. Therefore, a TCP receiver ack'ing more than 2
segments at a time causes the sending TCP to generate a larger
burst of traffic upon receipt of the ACK. This large burst of
traffic can overwhelm an intervening gateway, leading to higher
drop rates for both the connection and other connections passing
through the congested gateway.

In addition, the TCP slow start algorithm increases the congestion
window by 1 segment for each ACK received. Therefore, increasing
the ACK interval (thus decreasing the rate at which ACKs are
transmitted) increases the amount of time it takes slow start to
increase the congestion window to an appropriate operating point,
and the connection consequently suffers from reduced performance.
This is especially true for connections using large windows.

Relevant RFCs

RFC 1122
outlines delayed ACKs as a recommended mechanism.




Paxson, et. al. Informational [Page 41]

RFC 2525 TCP Implementation Problems March 1999



Trace file demonstrating it
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator). The advertised window (which never changed) and
timestamp options have been omitted for clarity, except for the
first packet sent by A:

12:09:24.820187 A.1174 > B.3999: . 2049:3497(1448) ack 1
win 33580 <nop,nop,timestamp 2249877 2249914> [tos 0x8]
12:09:24.824147 A.1174 > B.3999: . 3497:4945(1448) ack 1
12:09:24.832034 A.1174 > B.3999: . 4945:6393(1448) ack 1
12:09:24.832222 B.3999 > A.1174: . ack 6393
12:09:24.934837 A.1174 > B.3999: . 6393:7841(1448) ack 1
12:09:24.942721 A.1174 > B.3999: . 7841:9289(1448) ack 1
12:09:24.950605 A.1174 > B.3999: . 9289:10737(1448) ack 1
12:09:24.950797 B.3999 > A.1174: . ack 10737
12:09:24.958488 A.1174 > B.3999: . 10737:12185(1448) ack 1
12:09:25.052330 A.1174 > B.3999: . 12185:13633(1448) ack 1
12:09:25.060216 A.1174 > B.3999: . 13633:15081(1448) ack 1
12:09:25.060405 B.3999 > A.1174: . ack 15081

This portion of the trace clearly shows that the receiver (host B)
sends an ACK for every third full sized packet received. Further
investigation of this implementation found that the cause of the
increased ACK interval was the TCP options being used. The
implementation sent an ACK after it was holding 2*MSS worth of
unacknowledged data. In the above case, the MSS is 1460 bytes so
the receiver transmits an ACK after it is holding at least 2920
bytes of unacknowledged data. However, the length of the TCP
options being used [
RFC1323
] took 12 bytes away from the data
portion of each packet. This produced packets containing 1448
bytes of data. But the additional bytes used by the options in
the header were not taken into account when determining when to
trigger an ACK. Therefore, it took 3 data segments before the
data receiver was holding enough unacknowledged data (>= 2*MSS, or
2920 bytes in the above example) to transmit an ACK.

Trace file demonstrating correct behavior
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator), again with window and timestamp information
omitted except for the first packet:

12:06:53.627320 A.1172 > B.3999: . 1449:2897(1448) ack 1
win 33580 <nop,nop,timestamp 2249575 2249612> [tos 0x8]
12:06:53.634773 A.1172 > B.3999: . 2897:4345(1448) ack 1
12:06:53.634961 B.3999 > A.1172: . ack 4345
12:06:53.737326 A.1172 > B.3999: . 4345:5793(1448) ack 1
12:06:53.744401 A.1172 > B.3999: . 5793:7241(1448) ack 1
12:06:53.744592 B.3999 > A.1172: . ack 7241




Paxson, et. al. Informational [Page 42]

RFC 2525 TCP Implementation Problems March 1999



12:06:53.752287 A.1172 > B.3999: . 7241:8689(1448) ack 1
12:06:53.847332 A.1172 > B.3999: . 8689:10137(1448) ack 1
12:06:53.847525 B.3999 > A.1172: . ack 10137

This trace shows the TCP receiver (host B) ack'ing every second
full-sized packet, according to [
RFC1122
]. This is the same
implementation shown above, with slight modifications that allow
the receiver to take the length of the options into account when
deciding when to transmit an ACK."

So I guess the point is that at the rates we are discussing (the the according short periods between non-filtered ACKs the time-out issue will be moot). The Slow start issue might also be moot if the sender does more than simple ACK counting. This leaves redundancy... The fact that GRO/GSO effectively lead to ack stretching already the disadvantages might not be as bad today (for high bandwidth flows) than they were in the past...
Post by Mikael Abrahamsson
What benefit is there to either end system to send 35kPPS of ACKs in order to facilitate a 100 megabyte/s of TCP transfer?
Sounds like a lot of useless interrupts and handling by the stack, apart from offloading it to the NIC to do a lot of handling of these mostly useless packets so the CPU doesn't have to do it.
Why isn't 1kPPS of ACKs sufficient for most usecases?
This is not going to fly, as far as I can tell the ACK rate needs to be high enough so that its inverse does not exceed the period that is equivalent to the calculated RTO, so the ACK rate needs to scale with the RTT of a connection.

But I do not claim to be an expert here, I just had a look at some RFCs that might or might not be outdated already...

Best Regards
Sebastian
Post by Mikael Abrahamsson
--
Benjamin Cronce
2017-12-12 19:27:55 UTC
Permalink
Raw Message
Post by Sebastian Moeller
Hi Mikael,
Post by Mikael Abrahamsson
Post by Sebastian Moeller
Well, ACK filtering/thinning is a simple trade-off: redundancy versus
bandwidth. Since the RFCs say a receiver should acknoledge every second
full MSS I think the decision whether to filter or not should be kept to
Post by Mikael Abrahamsson
Why does it say to do this?
"2.13.
Name of Problem
Stretch ACK violation
Paxson, et. al. Informational [Page 40]
RFC 2525 TCP Implementation Problems March 1999
Classification
Congestion Control/Performance
Description
To improve efficiency (both computer and network) a data receiver
may refrain from sending an ACK for each incoming segment,
according to [
RFC1122
]. However, an ACK should not be delayed an
inordinate amount of time. Specifically, ACKs SHOULD be sent for
every second full-sized segment that arrives. If a second full-
sized segment does not arrive within a given timeout (of no more
than 0.5 seconds), an ACK should be transmitted, according to
[
RFC1122
]. A TCP receiver which does not generate an ACK for
every second full-sized segment exhibits a "Stretch ACK
Violation".
Significance
TCP receivers exhibiting this behavior will cause TCP senders to
generate burstier traffic, which can degrade performance in
congested environments. In addition, generating fewer ACKs
increases the amount of time needed by the slow start algorithm to
open the congestion window to an appropriate point, which
diminishes performance in environments with large bandwidth-delay
products. Finally, generating fewer ACKs may cause needless
retransmission timeouts in lossy environments, as it increases the
possibility that an entire window of ACKs is lost, forcing a
retransmission timeout.
It is interesting that enough of an issue occurred for them to explicitly
state that at least 1 ACK per 2 segments as an RFC. That being said, all
rules are meant to be broken, but not taken lightly when breaking. In
highly asymmetric connections with large bufferbloat, the sender is either
theoretically or practically of sending ACKs fast enough due to lack of
bandwidth, results in ACKs becoming highly delayed, which, in my opinion,
is worse. If the recover cannot ACK the receiver data within ~1.5 seconds,
the sender will resend the missing segments. In my experience, I have seen
upwards of 50% dup packet rates even though the actual loss rate was less
than 1%.

I do not feel that thinning ACKs gains much for any healthy ratio of
down:up. The overhead of those "wasteful" ACKs are on par with the overhead
of IP+TCP headers. Anything that can disturb the health of the Internet
should make strong measures to prevent the end user from configuring the
shaper in a knowingly destructive way. Like possibly letting the end user
configure the amount of bandwidth ACKs get. I see many saying 35k pps is
ridiculous, but that's pittance. If someone's network can't handle that,
maybe they need a special TCP proxy. Thinning ACKs to help with bufferbloat
is one thing, thinning ACKs because we feel TCP is too aggressive, is a can
of worms. Research on the topic is still appreciated, but we should be
careful about how much functionality Cake will have.
Post by Sebastian Moeller
Implications
When not in loss recovery, every ACK received by a TCP sender
triggers the transmission of new data segments. The burst size is
determined by the number of previously unacknowledged segments
each ACK covers. Therefore, a TCP receiver ack'ing more than 2
segments at a time causes the sending TCP to generate a larger
burst of traffic upon receipt of the ACK. This large burst of
traffic can overwhelm an intervening gateway, leading to higher
drop rates for both the connection and other connections passing
through the congested gateway.
In addition, the TCP slow start algorithm increases the congestion
window by 1 segment for each ACK received. Therefore, increasing
the ACK interval (thus decreasing the rate at which ACKs are
transmitted) increases the amount of time it takes slow start to
increase the congestion window to an appropriate operating point,
and the connection consequently suffers from reduced performance.
This is especially true for connections using large windows.
Relevant RFCs
RFC 1122
outlines delayed ACKs as a recommended mechanism.
Paxson, et. al. Informational [Page 41]
RFC 2525 TCP Implementation Problems March 1999
Trace file demonstrating it
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator). The advertised window (which never changed) and
timestamp options have been omitted for clarity, except for the
12:09:24.820187 A.1174 > B.3999: . 2049:3497(1448) ack 1
win 33580 <nop,nop,timestamp 2249877 2249914> [tos 0x8]
12:09:24.824147 A.1174 > B.3999: . 3497:4945(1448) ack 1
12:09:24.832034 A.1174 > B.3999: . 4945:6393(1448) ack 1
12:09:24.832222 B.3999 > A.1174: . ack 6393
12:09:24.934837 A.1174 > B.3999: . 6393:7841(1448) ack 1
12:09:24.942721 A.1174 > B.3999: . 7841:9289(1448) ack 1
12:09:24.950605 A.1174 > B.3999: . 9289:10737(1448) ack 1
12:09:24.950797 B.3999 > A.1174: . ack 10737
12:09:24.958488 A.1174 > B.3999: . 10737:12185(1448) ack 1
12:09:25.052330 A.1174 > B.3999: . 12185:13633(1448) ack 1
12:09:25.060216 A.1174 > B.3999: . 13633:15081(1448) ack 1
12:09:25.060405 B.3999 > A.1174: . ack 15081
This portion of the trace clearly shows that the receiver (host B)
sends an ACK for every third full sized packet received. Further
investigation of this implementation found that the cause of the
increased ACK interval was the TCP options being used. The
implementation sent an ACK after it was holding 2*MSS worth of
unacknowledged data. In the above case, the MSS is 1460 bytes so
the receiver transmits an ACK after it is holding at least 2920
bytes of unacknowledged data. However, the length of the TCP
options being used [
RFC1323
] took 12 bytes away from the data
portion of each packet. This produced packets containing 1448
bytes of data. But the additional bytes used by the options in
the header were not taken into account when determining when to
trigger an ACK. Therefore, it took 3 data segments before the
data receiver was holding enough unacknowledged data (>= 2*MSS, or
2920 bytes in the above example) to transmit an ACK.
Trace file demonstrating correct behavior
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator), again with window and timestamp information
12:06:53.627320 A.1172 > B.3999: . 1449:2897(1448) ack 1
win 33580 <nop,nop,timestamp 2249575 2249612> [tos 0x8]
12:06:53.634773 A.1172 > B.3999: . 2897:4345(1448) ack 1
12:06:53.634961 B.3999 > A.1172: . ack 4345
12:06:53.737326 A.1172 > B.3999: . 4345:5793(1448) ack 1
12:06:53.744401 A.1172 > B.3999: . 5793:7241(1448) ack 1
12:06:53.744592 B.3999 > A.1172: . ack 7241
Paxson, et. al. Informational [Page 42]
RFC 2525 TCP Implementation Problems March 1999
12:06:53.752287 A.1172 > B.3999: . 7241:8689(1448) ack 1
12:06:53.847332 A.1172 > B.3999: . 8689:10137(1448) ack 1
12:06:53.847525 B.3999 > A.1172: . ack 10137
This trace shows the TCP receiver (host B) ack'ing every second
full-sized packet, according to [
RFC1122
]. This is the same
implementation shown above, with slight modifications that allow
the receiver to take the length of the options into account when
deciding when to transmit an ACK."
So I guess the point is that at the rates we are discussing (the the
according short periods between non-filtered ACKs the time-out issue will
be moot). The Slow start issue might also be moot if the sender does more
than simple ACK counting. This leaves redundancy... The fact that GRO/GSO
effectively lead to ack stretching already the disadvantages might not be
as bad today (for high bandwidth flows) than they were in the past...
Post by Mikael Abrahamsson
What benefit is there to either end system to send 35kPPS of ACKs in
order to facilitate a 100 megabyte/s of TCP transfer?
Post by Mikael Abrahamsson
Sounds like a lot of useless interrupts and handling by the stack, apart
from offloading it to the NIC to do a lot of handling of these mostly
useless packets so the CPU doesn't have to do it.
Post by Mikael Abrahamsson
Why isn't 1kPPS of ACKs sufficient for most usecases?
This is not going to fly, as far as I can tell the ACK rate needs
to be high enough so that its inverse does not exceed the period that is
equivalent to the calculated RTO, so the ACK rate needs to scale with the
RTT of a connection.
But I do not claim to be an expert here, I just had a look at some RFCs
that might or might not be outdated already...
Best Regards
Sebastian
Post by Mikael Abrahamsson
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Dave Taht
2017-12-12 20:04:45 UTC
Permalink
Raw Message
What kicked off this thread was my urge to get a little code review of

https://github.com/dtaht/sch_cake/blob/cobalt/sch_cake.c#L904

because it seems to work well in a variety of tests, and ever better
as your D/U ratio cracks 10/1.

There are a few things I don't like about the implementation:

0) It's cpu intensive. Not as bad as hashing three times, as cake,
can, but intensive.

1) It reparses the whole queue looking for further acks to take out.
This makes sense were this to be applied to a single fifo, but in
cake's 1024 queue set associative case, leveraging something like
oldskb->isack && oldskb->hash == skb->hash on the five tuple would be
faster.

2) stopping entirely on a ack-filterable miss, and just working on
replacing the last packet on the tail of the queue, is far more O(1).
(I'd kind of like a "ackstate" machine, perhaps added to the cb)

3) I already made an attempt to deprioritize bulk ack flows slightly.

4) How much parsing of sack is really necessary?

5) I'm a little unsure as to the right things to do for ECN-echo. 1
ECN-echo packet should always be sent...

That all said, the results are fascinating, and I can live with all
these issues for a first release in sch_cake to more people in the
lede project... and plan on trying to pull this functionality out more
generally over the next month or three - primarily as impairments to
netem.
Post by Benjamin Cronce
Post by Sebastian Moeller
Hi Mikael,
Post by Mikael Abrahamsson
Post by Sebastian Moeller
Well, ACK filtering/thinning is a simple trade-off: redundancy versus
bandwidth. Since the RFCs say a receiver should acknoledge every second full
MSS I think the decision whether to filter or not should be kept to
Why does it say to do this?
"2.13.
Name of Problem
Stretch ACK violation
Paxson, et. al. Informational [Page 40]
RFC 2525 TCP Implementation Problems March 1999
Classification
Congestion Control/Performance
Description
To improve efficiency (both computer and network) a data receiver
may refrain from sending an ACK for each incoming segment,
according to [
RFC1122
]. However, an ACK should not be delayed an
inordinate amount of time. Specifically, ACKs SHOULD be sent for
every second full-sized segment that arrives. If a second full-
sized segment does not arrive within a given timeout (of no more
than 0.5 seconds), an ACK should be transmitted, according to
[
RFC1122
]. A TCP receiver which does not generate an ACK for
every second full-sized segment exhibits a "Stretch ACK
Violation".
Significance
TCP receivers exhibiting this behavior will cause TCP senders to
generate burstier traffic, which can degrade performance in
congested environments. In addition, generating fewer ACKs
increases the amount of time needed by the slow start algorithm to
open the congestion window to an appropriate point, which
diminishes performance in environments with large bandwidth-delay
products. Finally, generating fewer ACKs may cause needless
retransmission timeouts in lossy environments, as it increases the
possibility that an entire window of ACKs is lost, forcing a
retransmission timeout.
It is interesting that enough of an issue occurred for them to explicitly
state that at least 1 ACK per 2 segments as an RFC. That being said, all
rules are meant to be broken, but not taken lightly when breaking. In highly
asymmetric connections with large bufferbloat, the sender is either
theoretically or practically of sending ACKs fast enough due to lack of
bandwidth, results in ACKs becoming highly delayed, which, in my opinion, is
worse. If the recover cannot ACK the receiver data within ~1.5 seconds, the
sender will resend the missing segments. In my experience, I have seen
upwards of 50% dup packet rates even though the actual loss rate was less
than 1%.
I too have seen some insane dup packet rates also, but that's a
failure on the input side mostly.
Post by Benjamin Cronce
I do not feel that thinning ACKs gains much for any healthy ratio of
down:up.
Define "healthy".
Post by Benjamin Cronce
The overhead of those "wasteful" ACKs are on par with the overhead
of IP+TCP headers. Anything that can disturb the health of the Internet
should make strong measures to prevent the end user from configuring the
shaper in a knowingly destructive way. Like possibly letting the end user
configure the amount of bandwidth ACKs get. I see many saying 35k pps is
ridiculous, but that's pittance.
I tend to agree that for longer RTTs (and in the context of a clean
sheet design for TCP!)
more than one ack per ms is excessive - more broadly, per TXOP, in the
case of wifi in its presently overly-reliable mac retransmission
layer.

I sometimes wish we had a substrate for "I'm going to send 17 packets
total on this flow, tell me if you got 'em", rather than the ack
clock.
Post by Benjamin Cronce
If someone's network can't handle that,
maybe they need a special TCP proxy. Thinning ACKs to help with bufferbloat
is one thing,
I wouldn't quite define it that way. Thinning acks to make room for
non-acks in an already debloated
environment or "applying a congestion control algorithm that applies
specifically to acks". The blog posting and graph here showed how slow
codel was at clearing room here:

http://blog.cerowrt.org/post/ack_filtering/
Post by Benjamin Cronce
thinning ACKs because we feel TCP is too aggressive, is a can
of worms. Research on the topic is still appreciated, but we should be
careful about how much functionality Cake will have.
good point. Despite deployment in a few places like riverbed and ubnt
and lede, sch_cake has not achieved
particularly high penetration elsewhere. It's still kind of a
convenient research vehicle, but I dearly wish the things I love about
it (the deficit scheduler, the per host fq stuff) had more users.
Post by Benjamin Cronce
Post by Sebastian Moeller
Implications
When not in loss recovery, every ACK received by a TCP sender
triggers the transmission of new data segments. The burst size is
determined by the number of previously unacknowledged segments
each ACK covers. Therefore, a TCP receiver ack'ing more than 2
segments at a time causes the sending TCP to generate a larger
burst of traffic upon receipt of the ACK. This large burst of
traffic can overwhelm an intervening gateway, leading to higher
drop rates for both the connection and other connections passing
through the congested gateway.
This is no longer true in the case of pacing.
Post by Benjamin Cronce
Post by Sebastian Moeller
In addition, the TCP slow start algorithm increases the congestion
window by 1 segment for each ACK received. Therefore, increasing
the ACK interval (thus decreasing the rate at which ACKs are
transmitted) increases the amount of time it takes slow start to
increase the congestion window to an appropriate operating point,
and the connection consequently suffers from reduced performance.
This is especially true for connections using large windows.
Relevant RFCs
RFC 1122
outlines delayed ACKs as a recommended mechanism.
Paxson, et. al. Informational [Page 41]
RFC 2525 TCP Implementation Problems March 1999
Trace file demonstrating it
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator). The advertised window (which never changed) and
timestamp options have been omitted for clarity, except for the
12:09:24.820187 A.1174 > B.3999: . 2049:3497(1448) ack 1
win 33580 <nop,nop,timestamp 2249877 2249914> [tos 0x8]
12:09:24.824147 A.1174 > B.3999: . 3497:4945(1448) ack 1
12:09:24.832034 A.1174 > B.3999: . 4945:6393(1448) ack 1
12:09:24.832222 B.3999 > A.1174: . ack 6393
12:09:24.934837 A.1174 > B.3999: . 6393:7841(1448) ack 1
12:09:24.942721 A.1174 > B.3999: . 7841:9289(1448) ack 1
12:09:24.950605 A.1174 > B.3999: . 9289:10737(1448) ack 1
12:09:24.950797 B.3999 > A.1174: . ack 10737
12:09:24.958488 A.1174 > B.3999: . 10737:12185(1448) ack 1
12:09:25.052330 A.1174 > B.3999: . 12185:13633(1448) ack 1
12:09:25.060216 A.1174 > B.3999: . 13633:15081(1448) ack 1
12:09:25.060405 B.3999 > A.1174: . ack 15081
This portion of the trace clearly shows that the receiver (host B)
sends an ACK for every third full sized packet received. Further
investigation of this implementation found that the cause of the
increased ACK interval was the TCP options being used. The
implementation sent an ACK after it was holding 2*MSS worth of
unacknowledged data. In the above case, the MSS is 1460 bytes so
the receiver transmits an ACK after it is holding at least 2920
bytes of unacknowledged data. However, the length of the TCP
options being used [
RFC1323
] took 12 bytes away from the data
portion of each packet. This produced packets containing 1448
bytes of data. But the additional bytes used by the options in
the header were not taken into account when determining when to
trigger an ACK. Therefore, it took 3 data segments before the
data receiver was holding enough unacknowledged data (>= 2*MSS, or
2920 bytes in the above example) to transmit an ACK.
Trace file demonstrating correct behavior
Trace file taken using tcpdump at host B, the data receiver (and
ACK originator), again with window and timestamp information
12:06:53.627320 A.1172 > B.3999: . 1449:2897(1448) ack 1
win 33580 <nop,nop,timestamp 2249575 2249612> [tos 0x8]
12:06:53.634773 A.1172 > B.3999: . 2897:4345(1448) ack 1
12:06:53.634961 B.3999 > A.1172: . ack 4345
12:06:53.737326 A.1172 > B.3999: . 4345:5793(1448) ack 1
12:06:53.744401 A.1172 > B.3999: . 5793:7241(1448) ack 1
12:06:53.744592 B.3999 > A.1172: . ack 7241
Paxson, et. al. Informational [Page 42]
RFC 2525 TCP Implementation Problems March 1999
12:06:53.752287 A.1172 > B.3999: . 7241:8689(1448) ack 1
12:06:53.847332 A.1172 > B.3999: . 8689:10137(1448) ack 1
12:06:53.847525 B.3999 > A.1172: . ack 10137
This trace shows the TCP receiver (host B) ack'ing every second
full-sized packet, according to [
RFC1122
]. This is the same
implementation shown above, with slight modifications that allow
the receiver to take the length of the options into account when
deciding when to transmit an ACK."
So I guess the point is that at the rates we are discussing (the the
according short periods between non-filtered ACKs the time-out issue will be
moot). The Slow start issue might also be moot if the sender does more than
simple ACK counting. This leaves redundancy... The fact that GRO/GSO
effectively lead to ack stretching already the disadvantages might not be as
bad today (for high bandwidth flows) than they were in the past...
Post by Mikael Abrahamsson
What benefit is there to either end system to send 35kPPS of ACKs in
order to facilitate a 100 megabyte/s of TCP transfer?
Sounds like a lot of useless interrupts and handling by the stack, apart
from offloading it to the NIC to do a lot of handling of these mostly
useless packets so the CPU doesn't have to do it.
Why isn't 1kPPS of ACKs sufficient for most usecases?
This is not going to fly, as far as I can tell the ACK rate needs
to be high enough so that its inverse does not exceed the period that is
equivalent to the calculated RTO, so the ACK rate needs to scale with the
RTT of a connection.
But I do not claim to be an expert here, I just had a look at some RFCs
that might or might not be outdated already...
Best Regards
Sebastian
Post by Mikael Abrahamsson
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
David Lang
2017-12-12 21:03:04 UTC
Permalink
Raw Message
Post by Benjamin Cronce
I do not feel that thinning ACKs gains much for any healthy ratio of
down:up. The overhead of those "wasteful" ACKs are on par with the overhead
of IP+TCP headers.
assuming that there was no traffic going the other way to compete with the acks.
Post by Benjamin Cronce
Anything that can disturb the health of the Internet
should make strong measures to prevent the end user from configuring the
shaper in a knowingly destructive way. Like possibly letting the end user
configure the amount of bandwidth ACKs get. I see many saying 35k pps is
ridiculous, but that's pittance. If someone's network can't handle that,
maybe they need a special TCP proxy. Thinning ACKs to help with bufferbloat
is one thing, thinning ACKs because we feel TCP is too aggressive, is a can
of worms. Research on the topic is still appreciated, but we should be
careful about how much functionality Cake will have.
Yes, research is needed, but we need to recognize that what was appropriate when
1Mb was a very fast link may not be appropriate when you are orders of magnatude
faster, and where there can be significant amounts of traffic in the other
direction.

I think that TCP is pretty wasteful of bandwidth (and txops on wifi) under most
conditions.

Just chopping the number from 1/2 to 1/200 or something like that is obviously
wrong, but I have a real hard time figuring out how collapsing acks that are
sitting in a queue together into one ack will hurt. The acks that you are
deleting are not going to get to the recipient any faster than the ack that you
keep (at least if done correctly), so how can it make things better to delay
acking data that you have received in order to send out many additional acks of
parts of that data?

David Lang
Jonathan Morton
2017-12-12 21:29:39 UTC
Permalink
Raw Message
Taking into account a variety of scenarios, I have difficulty identifying a
case where an ack deleted by a reasonably conservative algorithm would have
given any practical benefit had it remained, *including* considerations of
smoothness of ack-clocking.

If the uplink isn't congested then no deletions occur; if it is congested
then there's a high probability that a flow-isolation scheme would deliver
several acks back to back between larger data packets, so an ack-clocked
sender would still be "lumpy". That's without even considering aggregation
and discrete MAC-grant links (ie. DOCSIS).

Deleting unnecessary acks from a congested uplink also frees capacity for
competing traffic, which I think we can agree is a good thing when it has
no deleterous side-effects.

I have not yet personally verified that the algorithm now in Cake matches
my assumptions. If it doesn't, I'll push for modifications.

Incidentally, I am of the opinion that ECE can safely be ignored for
ack-filtering purposes. Normally ECE remains set in all acks until a CWR
is heard in reply, so it only matters that the ECE signal isn't *delayed* -
which ack-filtering actually helps to achieve. More sophisticated uses of
ECE should also survive this as long as statistical independence is
maintained.

- Jonathan Morton
Jonathan Morton
2017-12-12 22:03:24 UTC
Permalink
Raw Message
The one "correct" argument against ack-filtering I've seen is that it
encourages (or rather validates) the use of extreme asymmetry ratios.

However, these extreme ratios are already in widespread use without the aid
of ack-filtering. Even ADSL2 Annex A has, as its "ideal" sync rate, a 16:1
ratio, which Annex M modifies to under 10:1. I fear we must conclude that
technical considerations are not the driving factor here.

A better place to sort out the asymmetry problem (as well as several other
problems of current interest) would be in a free, competitive market.
Sadly such a thing is rare in the telecoms sector.

- Jonathan Morton
David Lang
2017-12-12 22:21:39 UTC
Permalink
Raw Message
Post by Jonathan Morton
The one "correct" argument against ack-filtering I've seen is that it
encourages (or rather validates) the use of extreme asymmetry ratios.
I would sure rather have a extremely asymmetric ration than a 'proper' ratio
with the same upstream bandwidth.

I really doubt that choosing to badly support extreme ratios will stop or even
slow down the deployment, and the technology continues to develop in ways that
make such ratios more common (it's much easier to spend money on signal
processing for a central box than for all the endpoint products)

David Lang
Jonathan Morton
2017-12-12 22:41:44 UTC
Permalink
Raw Message
Actually, the cost argument goes the other way. You need heavy DSP to
*receive* high bandwidths; sending it is much easier computationally.
Also, in aggregate a hundred cheap CPE boxes probably have more DSP
horsepower than the one head-end box serving them.

What the centralised head-end has an advantage in is transmit power, and
thus SNR. This feeds into Shannon's equation and supports your argument
more directly. In ADSL this is partly compensated for by assigning the
lower frequency carriers to the upload direction, since they have less loss
than high frequencies on a copper pair.

However, in the most extreme examples I've seen, the level of asymmetry has
little to do with the underlying link technology and more to do with how
the provisioning was arbitrarily set up.

Things like ADSL with an unrestricted downlink sync rate but uplink limited
to 128k. Or DOCSIS with a huge headline bandwidth for downlink, and no
obvious mention (until you've paid for it, set it up and measured it) that
the uplink shaper is set to about a fiftieth of that. I seriously doubt
that DOCSIS is ever inherently that asymmetric.

- Jonathan Morton
Mikael Abrahamsson
2017-12-13 09:46:00 UTC
Permalink
Raw Message
Post by Jonathan Morton
the uplink shaper is set to about a fiftieth of that. I seriously doubt
that DOCSIS is ever inherently that asymmetric.
Well, the products are, because that's what the operators seems to want,
probably also because that's what the customers demand.

So my modem has 16x4 (16 downstream channels and 4 upstream channels),
meaning built into the hardware, I have 1/4 split.

Then providers typically (this is my understanding, I haven't worked
professionally with DOCSIS networks) do is they have 24 downstream
channels and 4 upstream channels. Older modems can have 8 downstream and 4
upstream for instance, so they'll "tune" to the amount of channels they
can, and then there is an on-demand scheduler that handles upstream and
downstream traffic.

So I guess theoretically the operator could (if large enough) make a hw
vendor create a 16x16 modem and have 32 channels total. But nobody does
that, because that doesn't sell as well as having more downstream (because
people don't seem to care about upstream). It just makes more market sense
to sell these asymmetric services, because typically people are eyeballs
and they don't need a lot of upstream bw (or think they need it).

On the ADSL side, I have seen 28/3 (28 down, 3 up) for annex-M with
proprietary extensions. The fastest symmetric I have seen is 4.6/4.6. So
if you as an operator can choose between selling a 28/3 or 4.6/4.6
service, what will you do? To consumers, it's 28/3 all day.

So people can blame the ISPs all day long, but there is still (as you
stated) physical limitations on capacity on RF spectrum in air/copper, and
you need to handle this reality somehow. If a lot of power is used
upstream then you'll get worse SNR for the downstream, meaning less
capacity overall. Symmetric access capacity costs real money and results
in less overall capacity unless it's on point to point fiber.
--
Mikael Abrahamsson email: ***@swm.pp.se
Jonathan Morton
2017-12-13 10:03:48 UTC
Permalink
Raw Message
Forgive my ignorance, but does each channel have the same capacity in both
directions in DOCSIS?

- Jonathan Morton
Sebastian Moeller
2017-12-13 12:11:15 UTC
Permalink
Raw Message
Forgive my ignorance, but does each channel have the same capacity in both directions in DOCSIS?
A quick look at https://en.wikipedia.org/wiki/DOCSIS seems to reveal that there typically is higher capacity for each downstream versus each upstream channel, at least downstream consistently seems to offer higher maximal modulations...

Best Regards
Sebastian
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Jonathan Morton
2017-12-13 12:18:55 UTC
Permalink
Raw Message
Okay, from the tables on that page, it seems that the most asymmetric
maximal configuration is below 8:1. That's in line with what you'd expect
given transmit power and thus SNR differences.

Hence no legitimate reason to provision at 42:1 and above...

- Jonathan Morton
Sebastian Moeller
2017-12-13 12:36:09 UTC
Permalink
Raw Message
Hi Mikael,
the uplink shaper is set to about a fiftieth of that. I seriously doubt that DOCSIS is ever inherently that asymmetric.
Well, the products are, because that's what the operators seems to want, probably also because that's what the customers demand.
Not 100% about the demand; I believe this also has a component of market segmentation:
a) everybody that actually wants to offer content is going to be not to well-served with the asymmetric links and hence might need to upgrade to the typical business-grade contracts that AFAIKT often have smaller download/upload ratios.
b) I seem to recall (but can find no evidence, so I might fantasizing) that having assymmetric traffic can have advantages for an ISP with peering/transit costs.
So my modem has 16x4 (16 downstream channels and 4 upstream channels), meaning built into the hardware, I have 1/4 split.
In addition to the differences in available modulations for down- and upstream channels.
Then providers typically (this is my understanding, I haven't worked professionally with DOCSIS networks) do is they have 24 downstream channels and 4 upstream channels. Older modems can have 8 downstream and 4 upstream for instance, so they'll "tune" to the amount of channels they can, and then there is an on-demand scheduler that handles upstream and downstream traffic.
So I guess theoretically the operator could (if large enough) make a hw vendor create a 16x16 modem and have 32 channels total.
But nobody does that, because that doesn't sell as well as having more downstream (because people don't seem to care about upstream).
Or because more symmetric offers can be sold for more money to businesses (sure the "business" contract class probably offers more than that, but I think this is one thing it does offer).
It just makes more market sense to sell these asymmetric services, because typically people are eyeballs and they don't need a lot of upstream bw (or think they need it).
Let's put it that way, people simply do not know as in the advertisements one typically only sees the downstream numbers with the upstream relegated to the footnotes (or hidden behind a link). If customers truly would not care ISPs could afford to be more open with the upstream numbers (something regulators would certainly prefer to hiding the information in the fine print).
On the ADSL side, I have seen 28/3 (28 down, 3 up) for annex-M with proprietary extensions. The fastest symmetric I have seen is 4.6/4.6. So if you as an operator can choose between selling a 28/3 or 4.6/4.6 service, what will you do? To consumers, it's 28/3 all day.
I agree that most users would see it that way (especially since 4.6 to 3 is not that much loss); also I b;eive it will be hard to offer simultaneous 23/3 and 4.6/4.6 over the same trunk line (not sure whether that is the correct word, I mean the thick copper cable "tree" that starts from the CO/gf-attached DSLAM).
For ADSL the challenge is that the up-/downstrewam bands need to be equal for all users on a trunk cable other wise interference/cross talk will be bad; and the most remote customer will still need some downstream effectively limiting the high end for the single upstream band in ADSL. VDSL2 sidesteps this issue somewhat by using multiple upstream bands and more remote lines will simply miss out on the higher frequency upstream bands but will still get a better symmetry...
So people can blame the ISPs all day long, but there is still (as you stated) physical limitations on capacity on RF spectrum in air/copper,
These limitations might or might not be close: https://www.assia-inc.com/wp-content/uploads/2017/05/TDSL-presentation.pdf
and you need to handle this reality somehow. If a lot of power is used upstream then you'll get worse SNR for the downstream, meaning less capacity overall. Symmetric access capacity costs real money and results in less overall capacity unless it's on point to point fiber.
Best Regards
Sebastian
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Luca Muscariello
2017-12-13 12:39:57 UTC
Permalink
Raw Message
If I understand the patch well, the ack filter is actually fixing the
problem of ACK compression only.
Because it is enforced on packets in the queue only. It is stateless.

ACK compression would happen even w/o highly asymmetric access links by
just
having concurrent data streams with ack streams.

So, IMO, the patch is harmless per se in all cases.
ACK compression is harmful though and the patch fixes it.

Background on ACK compression:

Lixia Zhang, Scott Shenker, and Daivd D. Clark.
Observations on the dynamics of a congestion control algorithm: the effects
of two-way traffic.
acm sigcomm 1991.
Post by Jonathan Morton
The one "correct" argument against ack-filtering I've seen is that it
encourages (or rather validates) the use of extreme asymmetry ratios.
However, these extreme ratios are already in widespread use without the
aid of ack-filtering. Even ADSL2 Annex A has, as its "ideal" sync rate, a
16:1 ratio, which Annex M modifies to under 10:1. I fear we must conclude
that technical considerations are not the driving factor here.
A better place to sort out the asymmetry problem (as well as several other
problems of current interest) would be in a free, competitive market.
Sadly such a thing is rare in the telecoms sector.
- Jonathan Morton
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Dave Taht
2017-11-29 18:41:41 UTC
Permalink
Raw Message
Post by Mikael Abrahamsson
The better solution would of course be to have the TCP peeps change the
way TCP works so that it sends fewer ACKs.
Which tends to perturb the way the TCP self-clocking feedback loop works,
and to break Nagle.
Linux TCP is no longer particularly ack-clocked. In the post pacing,
post sch_fq world, packets are released (currently) on a 1ms schedule.
Support was recently released for modifying that schedule on a per
driver basis, which turns out to be helpful for wifi.

see: https://www.spinics.net/lists/netdev/msg466312.html
Post by Mikael Abrahamsson
In the TCP implementations I tcpdump regularily, it seems they send one
ACK per 2 downstream packets.
That's the delack algorithm. One of the stupidest algorithms I've had the
displeasure of looking at (a fixed 500ms timeout, sheesh).
Nagle would probably agree. He once told me he wished for 1 ack per
data packet...

We were young then.
And yes, it breaks Nagle.
Post by Mikael Abrahamsson
I don't want middle boxes making "smart" decisions
Ironically, it was dave reed's (co-author of the end to end argument)
50x1 ratio network connection that was an impetus to look harder at
this, and what I modeled in
http://blog.cerowrt.org/post/ack_filtering/

(I note there is discussion and way more tests landing on the cake mailing list)

The astounding number was that we were able to drop 70% of all packets
(and 90+% of acks) without doing any visible harm on the tests.
I agree, especially if they use transport-layer data to make their
decisions.
I'm not particularly fond of the idea myself! But I didn't invent
severe network asymmetry, or cpus that can't context switch worth a
damn.
Post by Mikael Abrahamsson
Since this ACK reduction is done on probably hundreds of millions of
fixed-line subscriber lines today,
What I'd started with was wanting to create impairments for netem that
matched common ack-filtering schemes in the field already.
what arguments do designers of TCP have
Post by Mikael Abrahamsson
to keep sending one ACK per 2 received TCP packets?
this would be a good list to have.

I note osx does stretch acks by default.
I think it's about growing the TCP congestion window fast enough. Recall
that that AIMD counts received ACKs, not ACKed bytes.
the cake code has a specific optimization to preserve slow start. It
can be improved.
(And not breaking Nagle.)
-- Juliusz
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Steinar H. Gunderson
2017-11-29 23:29:03 UTC
Permalink
Raw Message
Post by Dave Taht
Linux TCP is no longer particularly ack-clocked. In the post pacing,
post sch_fq world, packets are released (currently) on a 1ms schedule.
Most Linux hosts don't run sch_fq, though. I mean, it's not even default in a
4.15 kernel.

/* Steinar */
--
Homepage: https://www.sesse.net/
Stephen Hemminger
2017-11-29 23:59:17 UTC
Permalink
Raw Message
On Wed, 29 Nov 2017 10:41:41 -0800
Post by Dave Taht
Post by Mikael Abrahamsson
The better solution would of course be to have the TCP peeps change the
way TCP works so that it sends fewer ACKs.
Which tends to perturb the way the TCP self-clocking feedback loop works,
and to break Nagle.
Linux TCP is no longer particularly ack-clocked. In the post pacing,
post sch_fq world, packets are released (currently) on a 1ms schedule.
Support was recently released for modifying that schedule on a per
driver basis, which turns out to be helpful for wifi.
see: https://www.spinics.net/lists/netdev/msg466312.html
Also TCP BBR has lost its initial luster since it is unfair and ignores
losses and ECN (see recent netdev paper).
Eric Dumazet
2017-11-30 00:21:24 UTC
Permalink
Raw Message
Post by Stephen Hemminger
On Wed, 29 Nov 2017 10:41:41 -0800
Post by Dave Taht
Post by Mikael Abrahamsson
The better solution would of course be to have the TCP peeps change the
way TCP works so that it sends fewer ACKs.  
Which tends to perturb the way the TCP self-clocking feedback loop works,
and to break Nagle.  
Linux TCP is no longer particularly ack-clocked. In the post
pacing,
post sch_fq world, packets are released (currently) on a 1ms
schedule.
Support was recently released for modifying that schedule on a per
driver basis, which turns out to be helpful for wifi.
see: https://www.spinics.net/lists/netdev/msg466312.html
Also TCP BBR has lost its initial luster since it is unfair and ignores
losses and ECN (see recent netdev paper).
Recent netdev paper (from Larry) mentioned that fq_codel is used.

fq_codel is stochastic, so not a fairness champion with many flows.
There is a reason we use fq [1] instead ;)

We asked Larry how to reproduce his (surprising) results, because we
suspect some setup error or bias. He has to update his github trees.

netem can be tricky to use properly.

[1] Although the choice of packet scheduler is no longer an issue with
BBR now TCP can fallback to internal pacing implementation.

About ECN : We do not enable ECN for edge communications, so BBR runs
without ECN being negotiated/accepted.

We will probably take care of this point soon, but we had more urgent
problems.
Benjamin Cronce
2017-12-11 20:15:51 UTC
Permalink
Raw Message
I wonder if TCP could be effectively changed to send an ACK every
WindowSize/N number of packets. We'd need to be careful about how this
would affect 'slow start'.
Post by Dave Taht
Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.
Post by Dave Taht
http://blog.cerowrt.org/post/ack_filtering/
And quite noticeable at 16x1 ratios as well.
I'd rather like to have a compelling list of reasons why not to do
this! And ways to do it better, if not. The relevant code is hovering
https://github.com/dtaht/sch_cake/blob/cobalt/sch_cake.c#L902
Your post is already quite comprehensive when it comes to downsides.
The better solution would of course be to have the TCP peeps change the
way TCP works so that it sends fewer ACKs. I don't want middle boxes making
"smart" decisions when the proper solution is for both end TCP speakers to
do less work by sending fewer ACKs. In the TCP implementations I tcpdump
regularily, it seems they send one ACK per 2 downstream packets.
At 1 gigabit/s that's in the order of 35k pps of ACKs (100 megabyte/s
divided by 1440 divided by 2). That's in my opinion completely ludicrous
rate of ACKs for no good reason.
I don't know what the formula should be, but it sounds like the ACK
sending ratio should be influenced by how many in-flight ACKs there might
be. Is there any reason to have more than 100 ACKs in flight at any given
time? 500? 1000?
My DOCSIS connection (inferred through observation) seems to run on 1ms
upstream time slots, and my modem will delete contigous ACKs at 16 or 32
ACK intervals, ending up running at typically 1-2 ACKs per 1ms time slot.
This cuts down the ACK rate when I do 250 megabit/s downloads from 5-8
megabit/s to 400 kilobit/s of used upstream bw.
Since this ACK reduction is done on probably hundreds of millions of
fixed-line subscriber lines today, what arguments do designers of TCP have
to keep sending one ACK per 2 received TCP packets?
--
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
Dave Taht
2017-11-29 18:48:48 UTC
Permalink
Raw Message
Post by Dave Taht
Recently Ryan Mounce added ack filtering cabilities to the cake qdisc.
If I read this posting right, you're only measuring bulk performance.
What about interactive traffic, when there's only one or two data segments in
flight at a given time
In this design, you can only filter out an ack when you have a queue of them.

I am thinking saying "filter" has been misleading. Tho plenty
stateless ack filters exist.

ack-queue-compression?
Post by Dave Taht
I'd rather like to have a compelling list of reasons why not to do
this!
I haven't looked at Cake in detail, and I haven't put much thought into
- not risking breaking any of the TCP-related algorithms that depend on
ACKs arriving in a timely manner (AIMD, Nagle, Eifel, etc.),
especially in the case of just one segment in flight;
- not contributing to the ossification of the Internet by giving an
unfair advantage to TCP over other protocols;
- limiting the amount of knowledge that middleboxes have of the
transport-layer protocols, which leads to further ossification;
- avoiding complexity in middleboxes, which leads to a more brittle
Internet;
- not encouraging ISPs to deploy highly asymmetric links.
I'll add these to my list!
This is not my area of expertise, and therefore I don't feel competent to
have an opinion, but I think that before you deploy ACK filtering, you
really should consider the worries expressed above and whatever other
worries more competent people might have.
been worrying ever since I touched the wet paint!
-- Juliusz
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Jonathan Morton
2017-12-11 18:30:55 UTC
Permalink
Raw Message
While I haven't yet studied the ack filtering code in depth, Ryan indicated
that some of those concerns were considered in its design.

In particular, it won't ever delete the last remaining ack in a flow's
queue, only replace it with a more recently arrived one. That should take
care of interactive performance, which has always been high on Cake's
design considerations.

Cake itself is also careful to not drop the last packet (of any type) in a
flow's queue, for similar reasons.

- Jonathan Morton
Loading...