Concurrent access at nfs version 3

[okyou]

Hi all,

Our nfs server running on 11.11 system which share a filesystem to hundreds of clients.in most time it works well but sometimes it very slow to all our client to access nfsserver in busy time.my question is if there a parameter like NFSD_MAX_CONNECTION in nfs version4 which define the concurrent access for nfs version 3?
thanks alots:)

[Kapil Jha]

man nfs
http://compute.cnr.berkeley.edu/cgi-bin/man-cgi?nfs+4

*******************
NFSD_MAX_CONNECTIONS=num
Sets the maximum number of concurrent, connection- oriented connections. The default is unlimited and is obtained by not setting (that is, commenting out) NFSD_MAX_CONNECTIONS.Equivalent to the -c option in nfsd.
*******************

If there is a lot of netowrk traffic, can you please check if resource utilization.

BR,
Kapil+

[okyou]

sorry for my bad english. our nfs version now is 3,no NFSD_MAX_CONNECTION parameter

[Kapil Jha]

Can you please let us know below parameter value from /etc/rc.config.d/nfsconf

NUM_NFSD
NUM_NFSIOD

You may have to tune these paqrameters.

BR,
Kapil+

[okyou]

NUM_NFSD=16
NUM_NFSIOD=16

we have 16 cpus in this server. at it's busy time,abount 1600% cpu used.

[rick jones]

The slowness is then probably queueing waiting for service by an nfsd. I would expect idle CPU if there weren't enough concurrent TCP connections to enable the system to be driven to capacity. And 1600% CPU utilization on a 16 CPU system certainly seems like capacity.

If a great many of your operations are reads you might consider adding RAM to the system and/or allowing the filesystem buffer cache to be larger.

It might also be good to check things like netstat and lanadmin statistics.

You might also consider an upgrade to a more powerful system...

[okyou]

Thanks for your time.

we have 32GB total physical memory,1.2GB used for cache.i don't know if it is enough for our system. most time,our server only 2% cpu used and with over 20GB free memory left,it seems waste to upgrade to a more powerful system.any more suggestions?

[rick jones]

If the system is an NFS file server, why such a small filesystem buffer cache?

[Kapil Jha]

Since you are running huge NFS, I would recommend to increase the number of biod and nfsd as per me....approx 30-30 and check the performance.

If your buffer-cache is low and you are not running short of memory then you can increase this parameter it would help for sure.

BR,
Kapil+

[Dave Olker]

Are your NFS clients using TCP or UDP to access the NFS server? That will determine whether sizing the NUM_NFSD will help or not.

Regards,

Dave

[okyou]

Our server not only used for nfs but also for ldap server.and there is a bug in ldap cause the memory leak,so we leave a huge memory left for it to cross the working time:( may be it is a good sulution to set dbc_max to 20% when we got down time.

our client use siebel(tcp) to connect nfs server.when clerk use sieble at one time,nfsktcpd process use 1600% cpu,what's the benefit to set biod and nfsd to 30-30?

[Dave Olker]

Since you're using TCP, there will be no difference in setting NUM_NFSD or NUM_NFSIOD. Those are for the UDP NFS server daemons and the client-side biod daemons respectively - neither of which has anything to do with the NFS/TCP server side. You can forget about tuning those daemons.

Dave