[e2e] TCP performance vs Path MTU

[Vernon Schryver]

> From: "Oleg Vishnepolsky" <oleg at inforocket.com>

> Does anyone have measurements that show dependency of TCP
> performance (for bulk data transfers, as an example) on MTU ?
> Thanks... Oleg Vishnepolsky

Many of us have made such measurements mean times over the years.
It was easier when computers were slower and FDDI was more common.
You merely adjusted the MTU and ran your favorite throughput benchmark.

Today when almost all computers can saturate any LAN that most people have
available (i.e. 100 MHz Ethernet), it's harder.  Most people don't have
access to a medium that goes at GBytes/sec and supports MTU's from 500 to
at least several KByte.  Still, you can do the old measuresments with FDDI
by measuring the effects of a your TCP throughput benchmark on a "CPU
soaker."  That's a process that burns as many CPU cycles as it can get
as it counts them.  One possibility is a CPU benchmark that runs for 10
of minutes.  Your throughput benchmark will always report the medium
maximum for your 100 MBit/sec Ethernet, but as you vary the MTU, you'll
notice that the CPU soaker makes more or less progress.

The classic measurements over FDDI were approximately linear throughput
(or bytes transferred/CPU cycle) as a function of MTU after the MTU is
larger than 500 or 1000 bytes.  Below 500 bytes, throughput (or
bytes/cycle) varied more than linearly as the various fixed costs per
packet stopped decreasing with smaller packets.


Vernon Schryver    vjs at rhyolite.com