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

nfs performance tuning for hp-ux 11.0 and 11i systems page 14

Notes:

Page 14July 22, 2002

local

filesystems

Large Directory Dilemma

Customer Reported Problem

• V-Class system running HP-UX 11.0

• 12GB of physical memory

• Dynamic buffer cache (min=3% / max=10%)

Ø “ls -l” in NFS-mounted directory takes 30 minutes

to complete and consumes 98% of the CPU resources

Ø The same system using an 8MB static buffer cache

takes 90 seconds to complete the same command

Question: Why would a simple “ls -l” command issued in a large NFS-mounted

directory take a long time to complete and consume huge amounts of

system resources when a large buffer cache is configured on the client?

Answer: When reading an NFS-mounted directory, an rnode structure must be

allocated for every entry in the directory, and only a fixed number of

rnodes are available on the client (sized by nrnode, which in turn is

sized by ncsize or ninode). Once all free rnodes are used, the system

must begin reusing existing rnodes. In order to reuse an rnode, the

system must first invalidate all pages associated with the rnode from the

buffer cache. This requires the client to perform a serial search of the

entire buffer cache looking for pages to invalidate. When traversing a

directory containing many thousands of entries, the system will need to

search buffer cache many times. With a large buffer cache configured,

these serial searches can consume high amounts of CPU resources.

Fortunately, the buffer cache subsystem has be redesigned in HP-UX 11i

to eliminate this problem. Refer to page 108 for more information.