[e2e] High Packet Loss and TCP

[Rick Jones]

Greg Minshall wrote:

> i'm surprised by the 30% number as it seems high.  i guess my rule of
> thumb has been that TCP might behave itself under 10%, but that over
> 10% you're unlikely to get anything decent in terms of performance.

That is also my experience.

> maybe this is the difference between theory and practice (which
> difference is that in theory there is none, but in practice there is),
> or maybe the current turbo-charged TCPs (fast retransmit, etc.) really
> bump the performance in a lossy environment.

I've uploaded the slides from my portion of an HP-internal (though the
content is free of HP-internal only stuff) training course from years
ago (the document actually goes back to well before 1999, when it was
updated for ndd rather than nettune).  Long enough ago that the source
file is Interleaf, which I no longer have.  Thankfully, I still have the
.ps version:

ftp://ftp.cup.hp.com/dist/networking/briefs/SE_372_2_99.ps.Z

The 14th slide has a graph showing the throughput for a given window
size/MSS combination and average packet loss rate (IIRC, I hacked a
BSD_like ip_input.c to drop packets with a given value at a given offset
the first time it was seen, hacked netperf to put random values at that
offset).  By 10% packet loss throughput was basically epsilon.  The
primary intent of the slide was to demonstrate the benefits of fast
retransmit.  This is old enough that there is no SACK or anything, the
newest wizziest stuff is fast retrans.

rick jones
-- 
Wisdom Teeth are impacted, people are affected by the effects of events.
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to post, OR email to raj in cup.hp.com  but NOT BOTH...