[195138] in North American Network Operators' Group
Netflix fast.com performance
daemon@ATHENA.MIT.EDU (Amos Rosenboim)
Mon Jun 26 15:25:00 2017
X-Original-To: nanog@nanog.org
From: Amos Rosenboim <amos@oasis-tech.net>
To: "nanog@nanog.org" <nanog@nanog.org>
Date: Sun, 25 Jun 2017 19:22:59 +0300
Cc: Nitzan Tzelniker <nitzan.tz@oasis-tech.net>,
Shahar Grin <shaharg@oasis-tech.net>
Errors-To: nanog-bounces@nanog.org
Hello,
Lately we have been troubleshooting complaints from customers of several IS=
Ps about
relatively low results when testing to Netflix's fast.com
When we started troubleshooting we notice the following:
1. When the latency to the fast.com test server is ~70ms results are signif=
icantly lower than speedtest.net results or Google fiber results.
2. When latency to the fast.com server is low (10-20ms, over DSL link) the =
results are similar between speedtest.net and fast.com
When looking at packet captures of the fast.com test we notice the followin=
g:
1. fast.com pushes traffic very aggressively: we have applied a shaper with=
CIR=3D100Mbps and 200ms and even 500ms queue limit (cisco jargon)
and still noticed drops when performing fast.com tests.
2. Packets originated from fast.com do not have the tcp push (psh) flag on.
3. It seems that fast.com recovery from packet loss is not very good over h=
igh latency (~70ms).
It seems to us that after a packet loss even the server reaches his cwnd an=
d stops transmitting for relatively long periods.
We are interested to know if anybody else experience similar behaviour (wit=
h similar latency) or have any other insights into this.
If you would like to take a look at captures performed over 70ms latency pl=
ease PM and I'll share the files.
Regards
Amos
