Excellent NFS Performance Document

[Berlene Herren]

A document that details NFS performance issues and tuning was just released on the ITRC. Look for document ID KBAN00000261.

Berlene

[ISS Unix Admin]

In this document the following statement is made:-

The same 10%/300MB guideline holds true for an NFS Server using HFS disks. However, most NFS Servers us VxFS filesystems, which do not have these same buffer cache issues. Size the buffer cache with VxFS filesystems to whatever size will maximize HP-UX VxFS Performance (e.g. 50% or larger)

Does this sizing only apply to NFS servers or is it implying that when running VxFS (NFS or not) it is best to up the buffer cache to 50%?

I am trying to get a more definitive answer to buffer cache. I know the preferred method is to tweak dbc_max/min_pct and watch the CPU, buf read& write parameters from sar - but is there a starting figure one can use for VxFS file systems in a non-NFS environment?

Thanks


Leon

[CHRIS_ANORUO]

The europe support link for this document is at: http://europe-support.external.hp.com/cki/bin/doc.pl/sid=cc29109f003d5dee09/screen=ckiSearchResults

[Dan Hetzel]

Thanks Berlene !

Chris,
It seems that your link returns with a 'search string empty' message.
Maybe this one will work:
http://europe-support.external.hp.com/cki/bin/doc.pl/sid=7169ccd900193aeb84/screen=ckiSearchResults

Best regards,

Dan

[Dan Hetzel]

It doesn't work either 8-(

[Brian Hackley]

In response to Leon's question about buffer cache sizing for NFS environments, I can say we should be updating this document in about a month or so. Here is one planned update in the area of buffer cache:
There is no one right answer for every NFS client, however a good starting point for most clients is 400Mb or 25% of physical memory ? whichever is LESS.
We have also discovered that very large buffer cache on NFS servers does not always mix well with write-intensive workloads. The buffer cache on the server needs to be large enough to
allow the nfsd daemons to take write data off the network and post it to the cache,
to allow the daemons to reply quickly to the client. However, if the cache is too
large then brief periodic hangs can result when the syncer(1M) daemon flushes the
cache to disk. While there is no one right answer for every NFS server, a good
starting point for most servers is 1Gb or 50% of physical memory ? whichever is LESS.
Brian