*BSD News Article 89206


Return to BSD News archive

Path: euryale.cc.adfa.oz.au!newshost.carno.net.au!harbinger.cc.monash.edu.au!news.cs.su.oz.au!metro!metro!munnari.OZ.AU!news.ecn.uoknor.edu!news.wildstar.net!news.ececs.uc.edu!news.kei.com!news.mathworks.com!howland.erols.net!newspump.sol.net!newsfeeds.sol.net!hammer.uoregon.edu!zephyr.texoma.net!uunet!in3.uu.net!192.207.105.50!prodigy.com!darkstar.prodigy.com!davidsen
From: davidsen@tmr.com (bill davidsen)
Newsgroups: comp.os.linux.networking,comp.unix.bsd.bsdi.misc,comp.unix.bsd.misc
Subject: Re: Linux vs BSD
Date: 11 Feb 1997 21:10:04 GMT
Organization: TMR Associates, Schenectady NY
Lines: 32
Message-ID: <5dqn7c$459u@usenet1y.prodigy.net>
References: <32DFFEAB.7704@usa.net> <87d8uqu5vt.fsf@localhost.xs4all.nl> <5daqts$1f74@usenet1y.prodigy.net> <5dd8mp$la7@cynic.portal.ca>
NNTP-Posting-Host: darkstar.prodigy.com
Originator: davidsen@darkstar.prodigy.com
Xref: euryale.cc.adfa.oz.au comp.os.linux.networking:68496 comp.unix.bsd.bsdi.misc:6006 comp.unix.bsd.misc:2507

In article <5dd8mp$la7@cynic.portal.ca>,
Curt Sampson <cjs@cynic.portal.ca> wrote:
| In article <5daqts$1f74@usenet1y.prodigy.net>,
| bill davidsen <davidsen@tmr.com> wrote:
| >
| >In user or kernel space you should be able to drive the hardware to
| >it's limits, and that's what counts.
| 
| With NFS in user space you are in the kernel when you receive an
| NFS request. You must then leave the kernel to start processing
| the request, re-enter the kernel to do the disk I/O, return to
| userland to finish processing the request, and re-enter the kernel
| again to send the data. These transitions are expensive, and that
| is why an in-kernel implementation of NFS is theoretically more
| efficient than an equivalant user-land implementation of NFS.

I don't disagree that it can be more efficient, but to blame the
user mode nfs for poor performance on a dedicated server over 10Mbit
ethernet doesn't seem likely. Your technical point is totally
correct, but I think my comment about the use of user space being
inherently the cause of slow NFS, as was discussed in earlier posts,
is still valid; good user mode code should swamp the transport
medium. At the cost of more CPU, absolutely.

Any actual measurements tending to prove me wrong are welcome, the
only servers I can really load are on TokenRing, and I don't think
that's a good place to make the text.
-- 
bill davidsen (davidsen@tmr.com)
  Company policy prevents me from commenting on the performance of
this distributed database. However, the machines on which it runs
are called bottleneck and roadblock.