Making sense of NFS performance values...

[Monty Phillips]

I'm a bit new to NFS and in looking around, it looks like NFS performance values are typically reported in ops/sec. Is there a way to translate ops/sec to something like MB/s or kb/s? It doesn't look like it from what I have read.

But, the end result I am trying to determine is if a server will have enough throughput for 100 concurrent users writing a 10GB file via NFS. But, I don't know how to take these ops/sec values and correlate it with throughput.

Any help in helping me understand this would be much appreicated.

[Steven E. Protter]

That is a big meal for NFS to eat.

It is however possible.

Recommendations(partially borrowed from B. Hassel):

1) Keep the PC's of your network off the same collision domain as the NFS. Use VLAN to handle it. PC's can clog the network and slow down NFS traffic.

2) Set the NFS backend up on its own private network. It can be the same network as you boot Ignite, but you also want the fastest NIC on the job.

3) Gigabit networking between the HP-UX NFS client and NFS server.

SEP

[Monty Phillips]

Hi Steven,

Thanks for the info and advice. So, am I right to say that it is hard to translate ops/sec to MB/sec? Sounds like there are all kinds of variables that would affect this conversion.

Thanks again,
Monty

[David Nixon · Accepted solution]

First you need to know the percentage mix of call types generated by your applications -'nfsstat' would be the tool to use.

Estimating the size of NFS operations is not so easy: I looked at this a while back
for NFS version 2, under HP-UX 10.20.

Using information from 'tcpdumps' of NFS packets I estimated the mean size of a "gettatr" call as 324 bytes; compared to 8965 bytes for a "read".

You could make a rough estimate by allowing 8K - or whatever your NFS read/write sizes are set to - for read/write ops and say 300 bytes for other types...

[Monty Phillips]

David,

Thanks for the help and information. I appreciate it.

Monty