*BSD News Article 61776


Return to BSD News archive

Path: euryale.cc.adfa.oz.au!newshost.anu.edu.au!harbinger.cc.monash.edu.au!news.bhp.com.au!mel.dit.csiro.au!actcsiro!news.nsw.CSIRO.AU!metro!metro!inferno.mpx.com.au!news.mel.aone.net.au!imci4!newsfeed.internetmci.com!news.kei.com!nntp.coast.net!news00.sunet.se!sunic!news99.sunet.se!nntp-trd.UNINETT.no!due.unit.no!sthaug
From: sthaug@nethelp.no (Steinar Haug)
Newsgroups: comp.unix.bsd.freebsd.misc
Subject: Re: Glaring anomaly in lmbench results: socket bandwidth using localhost = 0.21 MB/s
Date: 11 Feb 1996 23:34:46 GMT
Organization: Nethelp Consulting, Trondheim, Norway
Lines: 19
Message-ID: <4fluem$rps@stork.runit.sintef.no>
References: <4fkus7$kab@stork.runit.sintef.no> <4fl75p$rlv@nntp5.u.washington.edu>
	<4fkuq0$1ev@dyson.iquest.net>
NNTP-Posting-Host: trane.uninett.no
In-reply-to: root@dyson.iquest.net's message of Sun, 11 Feb 1996 14:34:40 GMT

[John S. Dyson]

|   Actually, it is a good performer.  The effect that you are seeing is due
|   to an unfortunate clash between the size of the loopback mtu, some new
|   tcp code and an option set in the lmbench code.  The problem is that it
|   hasn't been clearly determined what to do about it.  Larry McVoy has
|   authorized a very minor mod to lmbench.  I think that is okay, but to
|   get an actual idea of the loopback performance without changing the
|   benchmark (loopback is what is being measured -- not TCP performance
|   with a network adapter), you might want to change the def for LOMTU
|   in /sys/net/if_loop.c from 16K to 4K.  You'll see nicer (and more
|   accurate reflection of the perf) numbers.

OK, I tried 4K, and got consistently 11.80 MByte/s. Much nicer, thanks! I
also tried 32K, which is the default in NetBSD-1.1, and that gave me an even
better loopback performance of 15.50 MByte/s. Now tell me -  why shouldn't I
leave the loopback MTU at 32K?

Steinar Haug, Nethelp consulting, sthaug@nethelp.no