[e2e] Re: TCP performance vs Path MTU

Nivedita Singhvi nivedita at sequent.com
Wed Apr 11 09:21:07 PDT 2001


> "Oleg Vishnepolsky" <oleg at inforocket.com>@postel.org on 04/11/2001 08:15:35
> AM

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


This is just a typical sample of what I'm getting on my boxes,
not any kind of formal benchmark.

Well, a very simple test that varies MTU between 256 and 1500
on a 100Mb lan between a 400MHz Pentium II and a 866MHz Pentium III
256 MB using a 3C905C 3COM ethernet adapter shows the following on
a linux 2.4.2 kernel:

Running a single netperf tcp stream test can saturate the link:

Socket Size (snd, rcv) = 131070 Bytes
Message Size           = 64 Bytes

MTU Size         Thruput       Confidence
(Bytes)          Mb/sec        %   # iter
-------          -------       ----------
256               68.85         99   15
300               73.11         99   15
450               81.74         99   15
512               83.58         99   15
536               84.29         99   15
650               86.85         99   15
800               89.23         99   15
1000              91.31         99   15
1200              92.71         99   15
1490              94.03         99   15
1500              94.05         99   15            

You normally wouldnt want to do this...

thanks,
Nivedita



More information about the end2end-interest mailing list