Re: [ccp4bb]: NFS Performance

["Kay Diederichs" <kay.diederichs@uni-konstanz.de>]

***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***

To Peter and all other contributors to this thread -

I would like to clarify that the NFS performance problem exists despite the
IRIX server running the latest 6.5.15m, and the Linux side being patched
with Myklebust's NFS patches (for 2.4.17). The problem is also not sensitive
to NFS2 or NFS3 mounts, and changing rsize, wsize does not help either. By
the way, with NFS3 rsize and wsize are as high as 32768 by default (look at
/proc/mounts).

I have been using bonnie and other benchmarks for years and I'm quite
satisfied with NFS performance and stability of both Linux and SGI. I find
that on a 100Mbit switched network, both read and write performance can
reach the upper limit given by the network speed, i.e. almost 10MB/s.

However, the problem shows when running SOLVE or some components of
arp/Warp. In both cases, CPU usage drops to below 1% during certain I/O
situations, and these long-running jobs simply don't advance. It appears as
if g77-compiled programs trigger some kind of drastic performance killer.

I believe that neither arp/Warp (I asked Victor about it) nor SOLVE (setting
SOLVETMPDIR to a local directory does _not_ help) honour CCP4_SCR -  if they
would the problem would probably not surface.

Maybe someone (Tassos?) could come up with a small piece of code that
exhibits the problem - in that case it would be possible to investigate. But
both programs are not open source so it's not easy to identify where the
problem arises.

Kay
----- Original Message -----
From: "Peter Brick" <p.brick@ic.ac.uk>
To: <ccp4bb@dl.ac.uk>
Sent: Wednesday, April 17, 2002 7:52 PM
Subject: RE: [ccp4bb]: NFS Performance


> ***  For details on how to be removed from this list visit the  ***
> ***          CCP4 home page http://www.ccp4.ac.uk         ***
>
> Hi
>
> >>
> >> check out the linux NFS kernel patches at
> >> http://www.fys.uio.no/~trondmy/src/2.4.18/
> >
> >Also note that there may be relevant patches on the Irix end:
> >
> >http://www.lib.uaa.alaska.edu/linux-kernel/archive/2002-Week-14/0765.html
> >
>
> This problem may be specific to SGI.
>
> I set up NFS using default values - and avoid patching Linux kernels if I
> can help it.
>
> I've used Bonnie to check NFS performance between Linux and Compaq Alpha
> machines.
>  RH7.2 with autofs and  TruUnix 5.1a automount  all with default
parameters.
>
>  Bonnie reports sequential (character putc()) o/p rates of:
>
> Local file
>
> Compaq: 23 MB/s
> Linux: 17.9 MB/s
>
> NFS
>
> Linux to Linux: 10.2 MB/s
> Compaq to Linux (scratch file):  9.9 MB/s
> Linux to Compaq (scratch file):  4.1 MB/s
>
> With 2GB scratch file but 1GB of RAM I can't avoid the effects of caching.
>
> Some of the numbers don't make much sense  - on my Alpha local block io
> gives a worse rate than char io while on a PC it is twice as fast.
>
> Peter
> --
> Peter Brick,
> Blackett Laboratory, Department of Biological Sciences,
> Imperial College, Prince Consort Road, London SW7 2BW, UK
> Tel: 020-7594-7704  Fax: 020-7589-0191
>
>
>
>