kernel hanging problem - nfs?...

[Wesley T. Perdue]

Jason,

Thanks again for your help.  I moved all clients to 1024-byte r/w blocks, and the whole-server hanging seems to have been resolved.

Individual processes are now hanging.  I increased the number of nfsds from 8 to 16; that seems to have helped, but I'll need to wait 'til tomorrow's heavy load for the final verdict.

I noticed in the current NFS-HOWTO:

http://nfs.sourceforge.net/nfs-howto/performance.html#MEMLIMITS

that they recommend increasing the socket read/write buffer size when increasing the number of nfs daemons.  Did you do that on your server(s)?

Thanks,
Wes

At 11:44 AM 12/28/2002 +1000, jason andrade wrote:
>On Thu, 26 Dec 2002, Wesley T. Perdue wrote:
>
>>
>> Nope; just switched 100.  I had read somewhere that 8192 was ideal on a LAN.  So 1024 is ideal for switched 100, based on the 1500-byte Ethernet packet and the 1024-byte disk block, correct?
>>
>
>8192 is great for a LAN provided you are not using a Linux NFS server with
>UDP (the default nfs server behaviour).  it works well if you have a solaris
>or *BSD or almost any other OS as the server and Linux as a client, but not
>any combination involving Linux as a NFS server.
>
>Linux does a reasonable job as a NFS server with 1K nfs packets.
>
>i believe the 1024 size has something to do with the way packet fragmentation
>is handled in linux and the interaction with the nfs implementation but i have
>heard various reasons and cannot claim to understand the problems in depth, only
>the symptoms (and some of the workarounds).
>
>regards,
>
>-jason