*BSD News Article 65385


Return to BSD News archive

Newsgroups: comp.unix.bsd.freebsd.misc
Path: euryale.cc.adfa.oz.au!newshost.anu.edu.au!newshost.telstra.net!act.news.telstra.net!vic.news.telstra.net!news.mira.net.au!news.vbc.net!alpha.sky.net!news.sprintlink.net!news.up.net!news.mtu.edu!newsxfer.itd.umich.edu!news.mathworks.com!gatech!swrinde!howland.reston.ans.net!vixen.cso.uiuc.edu!uwm.edu!news.inap.net!news1!not-for-mail
From: root@dyson.iquest.net (John S. Dyson)
Subject: Re: My recent comments on the Baker/ Lai paper at USENIX
X-Nntp-Posting-Host: dyson.iquest.net
Message-ID: <4kdoa2$nc@dyson.iquest.net>
Sender: news@iquest.net (News Admin)
Organization: John S. Dyson's Machine
References: <316999D7.167EB0E7@FreeBSD.org>
Date: Tue, 9 Apr 1996 13:22:10 GMT
Lines: 51

In article <316999D7.167EB0E7@FreeBSD.org>,
Jordan K. Hubbard <jkh@FreeBSD.org> wrote:
>
>I do think that benchmarking is important and that many types of useful
>"real world" data can be derived from them - our very own John Dyson
>puts significant amounts of time and effort into running benchmarks with
>Linux and other operating systems so as to "keep FreeBSD honest" in our
>own performance, looking for areas where they've made improvements which
>we have not.
>
I would like to re-iterate Jordan's comment in perhaps different words.  The
primary (and probably sole) purpose of my benchmarking is indeed to make sure
that we are not falling behind anywhere.  I do not normally disclose my
results except to people who ask (I have fed some info to Linus for example) or
need to know (perhaps once in a while I do post info just for general interest.)
The purpose is not marketing, and in fact the benchmark results are very
difficult to interpret.  The old adage "liars figure and figures lie" describes
a situation that I try to avoid, especially since the benchmarking is part of
the feedback mechanism that I use to make sure that the kernel and shared
library support developers aren't making things worse (and hopefully are
making things better)...  If I distort the results of my tests for marketing
reasons, it would destroy the advantage of doing the benchmarks for the FreeBSD
project.

A good example of something that is sorely lacking in existing benchmark
suites is that many of them are best run during idle system conditions.
Almost NO scalability under load is measured.  There are some very very
weak exceptions to this though (lat_ctx in lmbench, or the random seek
benchmark in bonnie.)

BTW, if anyone working on FreeBSD would like to submit ideas for macro-level
performance testing, let me know!!!  I will try to add them.  I generally
keep things quiet (for reasons that I have tried to describe above), but IMO
these measurements are an important part of the QC procedure on FreeBSD.

I don't have an organized, up-to-date set of results right now (I have been
comparing FreeBSD vs. NetBSD lately -- so I can't boot Linux today), but my
recent measurments show that FreeBSD is not falling behind in performance.  A
cute example of how bogus benchmarks misinform -- the lat_mmap benchmark
for FreeBSD vs. Linux shows that FreeBSD is very very much slower than
Linux...  Until one realizes that the cost of the FreeBSD mmap is very
quickly made up by the almost total elimination of pagefaults after the
mapping (esp. in the case of read-only segments, like .text).  There is an
aggregate performance increase by the tradeoff that we made.

The above is one reason that the adage previously mentioned applies to
some interpretations of certain benchmark results.

John Dyson
dyson@freebsd.org