300+ms of hotel wifi bufferbloat - peaking at 1.5 sec!

http://www.dslreports.com/speedtest/578850

I would get a kick out of it if folk here tried this new speedtest
periodically (on the "cable" setting) during the nanog conference. :wink:
There is a hires option for more detail on the resulting charts...

(or fiddled with "flent" (flent.org))

There's a corollary of the bufferbloat phenomenon: buffer drain time. It's not the size of the buffer, but how long it takes to empty. And US ISPs continue to say "customers don't want upload speed".
If the ISP upload speed was symmetric you'd likely never notice the 1-2MB of buffers.

I guess what I'm getting at is why do you continue to say buffers are too big instead of saying ISP upload is too slow?

While I agree that upload speeds aren't great, it doesn't mean that the buffers aren't big. Buffer sizes of the order of MB's are uncalled for at the edge, unless we're talking really high speeds. The miniscule performance increase for single TCP flows doesn't really justify the potential increase in latency for everyone else.

I did the dslreports tests on the NANOG wifi while listening to srikanth today:

http://www.dslreports.com/speedtest/593926

And my own (flent data also in this dir)...

http://snapon.lab.bufferbloat.net/~d/nanog/download_cdf.png

pretty good bandwidth. Pretty horrific latency... a couple detours
around the moon.

There's a bit of discussion on the AFMUG list about that speed test Dave. People with 500Mb, 1Gb,10Gb pipes were getting drastically different results depending on what "type" of test they did.

Josh Reynolds
CIO, SPITwSPOTS
www.spitwspots.com

There's a bit of discussion on the AFMUG list about that speed test Dave.
People with 500Mb, 1Gb,10Gb pipes were getting drastically different results
depending on what "type" of test they did.

There were also huge discussions of the dslreports testing ideas on
the bufferbloat "bloat" list.

Along the way we came up with some ideas and recommendations, which we
piled into a document here (comments welcomed!)

https://docs.google.com/document/d/1z5NN4WRKQKK-RtxtKR__XIwkybvsKEmunek2Ezdw_90/edit

I would like it very much if the currently named "fiber,cable,dsl,
etc" options in the dslreports test were renamed something like
"insane, extreme, medium, low" and let users call out a wifi or
wireless test vs

A big flaw in it as structured is that it first tests how robust a
network is to lots of flows in slow start in the early stages of the
test. The median idea = a grade needs work, also. But see above doc
and comment, please.

Also I have never trusted a browser to be able to drive tests like
this sanely, but I was pretty satisfied that the results I was getting
on the hardware I had (up to about 120Mbit) matched realities I could
also measure with the "flent.org" (formerly netperf-wrapper) tool. thx
for the steer to testers at higher rates.

Still, I would trust flent a LOT further than a browser at speeds
higher than that. It has been tested with reasonable results up to
40gigE.