NFS Performance Tuning for HP-UX 11.0 and 11i Systems

nfs performance tuning for hp-ux 11.0 and 11i systems page 47

Notes:

Page 47July 22, 2002

nfsd

What happens when an NFS request

arrives on the server?

• The kernel determines which per-CPU stream head to

queue the request on

• A single nfsd is awoken to handle the request

• CPUs can “task steal” – if an nfsd is ready to execute but

the CPU it is bound to is busy, a different CPU may

“steal” the waiting nfsd and execute it

• nfsds can “task steal” – an nfsd may steal requests from

other CPU’s queues

When an NFS request arrives on a multi-CPU server, the kernel must first determine

which CPU’s stream head to queue the request on. Remember that the server

creates separate pools of nfsds for each CPU and allocates an equal number of

daemons to each queue. Once the kernel decides where to queue the request, a

single nfsd is awoken to process the request.

If an nfsd is ready to execute but the CPU it is associated with is busy, an available

CPU may “steal” the nfsd and execute it. In addition, once an nfsd finishes

processing a request, before returning to sleep it will check the queue of the CPU it

is bound to to see if any other requests are waiting to be processed. If it finds a

request waiting it will service it. If the nfsd doesn’t find a request in the CPU’s

queue, the nfsd will then check the queues of the other CPUs to see if any requests

are waiting in their queues. If it finds a request waiting to be serviced, the nfsd

may “steal” the request from the queue of a different CPU and execute it. This is

commonly referred to as “task stealing.”