Discussion:
fremont flent node: what's the network setup there?
(too old to reply)
Aaron Wood
2017-10-05 05:41:44 UTC
Permalink
Raw Message
I'm comparing some numbers between the fremont node and a friend's Droplet
running netserver.

We've previous noted that we don't see more than a 120Mbps download rate
from the fremont node.

Today I was able to confirm in multiple back-to-back runs that the fremont
node was only giving me about 120Mbps of throughput, while both dslreports
and the droplet were giving me ~180Mbps (and 500ms of latency in the cable
head-end).

However, I noticed when running without SQM on my router here at home that
download latency from the fremont node was virtually non-existent, vs. the
very large latency that I was seeing from both the droplet and dsl-reports.

Box-plots:
https://drive.google.com/file/d/0B1xfqN4OiFEPODI4cEZjZ1NGeGVtc3RKbGJQR3ltaHhyb1Bz/view

That very, very low send latency from the fremont node makes me think that
it might be running BBR as the default congestion control algorithm? If I
run the cubic-reno test without any SQM, the results look pretty awful, but
without specifying the TCP cc alg, it seems to behave very, very well
(albeit not as much throughput as I think I should get).

Re-enabling cake (at 170Mbps downstream, 12Mbps up), I can't seem to get
much more than that 120Mbps from the router (even though I've tested it at
900Mbps as the netserver host itself). Am I running into a cpu limitation
in the NAT/forwarding modules?

-Aaron
Dave Taht
2017-10-05 06:39:25 UTC
Permalink
Raw Message
the simplest answer is for you to send me an ssh key I can put there.
second simplest is for me to fire up a new box and new kernel over
there and put your ssh key there also. It would not surprise me for
linode to have a rate limit somewhere elsewhere....

Lemme go spin up a new box in a minute.

It's currently sch_fq and cubic on kernel 4.8.6-x86_64-linode78.

also:

net.ipv4.tcp_ecn = 1
net.ipv4.tcp_ecn_fallback = 1
Post by Aaron Wood
I'm comparing some numbers between the fremont node and a friend's Droplet
running netserver.
We've previous noted that we don't see more than a 120Mbps download rate
from the fremont node.
Today I was able to confirm in multiple back-to-back runs that the fremont
node was only giving me about 120Mbps of throughput, while both dslreports
and the droplet were giving me ~180Mbps (and 500ms of latency in the cable
head-end).
However, I noticed when running without SQM on my router here at home that
download latency from the fremont node was virtually non-existent, vs. the
very large latency that I was seeing from both the droplet and dsl-reports.
https://drive.google.com/file/d/0B1xfqN4OiFEPODI4cEZjZ1NGeGVtc3RKbGJQR3ltaHhyb1Bz/view
That very, very low send latency from the fremont node makes me think that
it might be running BBR as the default congestion control algorithm? If I
run the cubic-reno test without any SQM, the results look pretty awful, but
without specifying the TCP cc alg, it seems to behave very, very well
(albeit not as much throughput as I think I should get).
Re-enabling cake (at 170Mbps downstream, 12Mbps up), I can't seem to get
much more than that 120Mbps from the router (even though I've tested it at
900Mbps as the netserver host itself). Am I running into a cpu limitation
in the NAT/forwarding modules?
-Aaron
_______________________________________________
Bloat mailing list
https://lists.bufferbloat.net/listinfo/bloat
--
Dave Täht
CEO, TekLibre, LLC
http://www.teklibre.com
Tel: 1-669-226-2619
Loading...