*BSD News Article 74219


Return to BSD News archive

Path: euryale.cc.adfa.oz.au!newshost.anu.edu.au!harbinger.cc.monash.edu.au!news.rmit.EDU.AU!news.unimelb.EDU.AU!munnari.OZ.AU!news.mel.connect.com.au!news.mira.net.au!vic.news.telstra.net!act.news.telstra.net!psgrain!iafrica.com!pipex-sa.net!plug.news.pipex.net!pipex!tank.news.pipex.net!pipex!news.mathworks.com!nntp.primenet.com!news.sprintlink.net!news-stk-3.sprintlink.net!news.sprintlink.net!news-ana-24.sprintlink.net!news.sprintlink.net!news-ana-7.sprintlink.net!hpg30a.csc.cuhk.hk!news.cuhk.edu.hk!newsf
eeder.ust.hk!nntp.hk.super.net!news.iij.ad.jp!news.CET.CO.JP!usenet
From: Michael Hancock <michaelh@cet.co.jp>
Newsgroups: comp.os.linux.networking,comp.unix.bsd.netbsd.misc,comp.unix.bsd.freebsd.misc
Subject: vfs overhead? (was Re: TCP latency)
Date: Sat, 20 Jul 1996 02:09:14 -0700
Organization: CET
Lines: 40
Message-ID: <31F0A23A.1019@cet.co.jp>
References: <4paedl$4bm@engnews2.Eng.Sun.COM> <31E80ACA.167EB0E7@dyson.iquest.net> <4sadde$qsv@linux.cs.Helsinki.FI> <31EA9FBC.41C67EA6@star-gate.com> <4sgo4l$ucf@linux.cs.Helsinki.FI>
NNTP-Posting-Host: a07m.cet.co.jp
Mime-Version: 1.0
Content-Type: text/plain; charset=iso-2022-jp
Content-Transfer-Encoding: 7bit
X-Mailer: Mozilla 2.02 (Win16; I)
Xref: euryale.cc.adfa.oz.au comp.os.linux.networking:45826 comp.unix.bsd.netbsd.misc:4120 comp.unix.bsd.freebsd.misc:24009

Linus Torvalds wrote:
> 
> Now, the factor of 3-4 doesn't necessarily tell the whole story, and it
> has been claimed that some of the overhead of FreeBSD is the VFS
> overhead.  Linux just seems to handle that same overhead a lot better,
> and WITHOUT needing to special case anything at all.

I hope this doesn't cause the backing out of the FS Stacking work put into system 
since 4.4Lite.

What can this do?

Stack a mirroring layer,
	a compression layer,
	a crypto layer,
	a union mount layer,
	
and do it on top of ffs,lfs,extfs, or whatever you come up with.

And you can do it without having to rewrite a lot of code.

This has extensible VOPs, etc.  Basically, there are some classical performance vs. 
(flexibility/extensibility) tradeoffs here and hope this benchmark war doesn't cause 
the FreeBSD camp to only look at the performance dimension.

The fs is one of the next great frontiers of development and the stuff that's in 
there is just too cool.  If there is going to be any optimizations, lets work within 
this fundamental framework.

Regarding the TCP/IP code there was talk about how the linear buffers in the Linux 
protocol stack might be a one of the reasons for the good numbers.  Van Jacobson has 
mentioned this in his works.

However, the performance gains don't seem to be that large over the linked list mbuf 
implementation in 4.4Lite.  I don't think it would be worth it to do the large 
contiguous buffer thing at this time, especially after looking at the Berkeley 
research post.  It would be a lot easier to merge in these developments if we don't 
make large changes to the existing system.

-mike hancock