[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-nfs
Subject:    Re: [NFS] Poor knfsd server performance on read() requests
From:       Neil Brown <neilb () cse ! unsw ! edu ! au>
Date:       2000-06-16 12:23:31
[Download RAW message or body]

On Thursday June 15, abz@document.co.uk wrote:
>                   Hi,
> 
>                   We have noticed very bad performance from our new
>                   nfs server. Our configuration is:
...
> 
>                   The symptoms are:
> 
>                   From either RedHat 5.2 or SunOS 4.1.4 or Solaris 2.5
>                   clients read performace is very, very slow. Write
>                   performace is quite good.
> 
>                   We have written a small benchmark program to test
>                   performace. The program issues a 'read()' request
>                   of 1024 bytes from the sample files on the nfs server.
> 
>                   The surprising thing is that the very first execution
>                   of the program is extremely slow - taking up to 30 secs
>                   to retreive 1Mb. Subsequent invocations on the same
>                   file are then typically sub 2 secs. We have run the
>                   same tests on SunOS, Solaris and RedHat 5.2 (Universal
>                   NFS Server 2.2beta37) nfs served files from the same
>                   varied clients listed above with no such dramatic
>                   delays for the first read. In fact the RedHat 5.2
>                   running on a lowly Dell Optiplex PIII produced nfs
>                   performance on par with a Sun Ultra 30.
> 
>                   There seems to be a really bad problem with read file
>                   access for files that are not in the nfs server
>                   cache.

I haven't noticed any problem with read throughput, so I don't think
that it is a fundamental problem with the NFS server.
It actually sounds more like a network problem.

Most other NFS servers will support tcp connections, and tcp
connections work better over a poor network.
Could you confirm if you are (probably by default) using tcp when
mounting from other servers, and udp when mounting from Linux/knfsd?
If that is the case, then I suspect that that is the determining
factor for the problem.
If that is not the case ... then I guess it must be something else.

Also, are you sure that when subsequent reads go much faster, that it
being served from the servers cache, and not from the clients cache? 

If you want to try, recent kernels like 2.4.0-test1-ac18 have
knfsd+tcp support as an experimental/developer option (it mostly works
but has some occasional problems and Servicee Denial
vulberabilities).  You could maybe try that to see if it gives better
service.

NeilBrown


> 
>                   Please could you help as this has significantly dented
>                   the image of Linux at our installation.
> 
>                   Thanks in advance,
> 
>                   AZ
> -- 
> smail       - Xerox Business Services Xerox (UK) Ltd. 
>               Document Imaging Services Centre, Beech House, 
>               Mitcheldean Gloucestershire GL17 0DD UK
> Phone	    - +(44) 1594 592797 Mobile - +(44) 966 477181
> www	    - www.document.co.uk  www home - www.rednet.co.uk/homepages/zydron
> 
> _______________________________________________
> NFS maillist  -  NFS@lists.sourceforge.net
> http://lists.sourceforge.net/mailman/listinfo/nfs

[prev in list] [next in list] [prev in thread] [next in thread]