The 'limit' of 4500 nfs/ops on the F630 concerns me in that combining the current nfs ops can easily put use near the upper limit of the box. What I don't have a feel for is how many nfs ops the current workload would actually put on a filer.

It is important for people to understand that LADDIS is a very hard load, and in real-world environments people will often see much higher throughput than LADDIS would seem to predict. I recently got a note from a customer "complaining" that his F230 was running at a sustained load of over 3300 ops per second, which is over double the LADDIS "limit" for that system.

How can that be? Here are some reasons:

(1) Some NFS operations are much harder than others. READ and WRITE, for instance, are hard, because they move data. GETATTR and LOOKUP are easy, because they do simple lookups.

LADDIS does 22% READs and 15% WRITEs. LADDIS does 50% of GETATTRs plus LOOKUPs. Run nfsstat to see what your current servers are doing. Most sites have fewer hard ops and more easy ones.

Also, look at the network load to your server. LADDIS generates about 1.5 MB of reads and 1 MB of writes per thousand LADDIS ops. So 4500 LADDIS ops represent a total of 11 MB/sec. (That's mega-bytes, not mega-bits.) If the network load to your server is lighter, then your load is probably lighter than LADDIS.

(2) LADDIS has a very poor cache-hit rate, so disk I/O becomes a bottleneck. (LADDIS goes out of it's way to access a very large working set.)

If your working set fits reasonably well in memory, your system will perform much better. Look at the cache age or cache hit statistics on your existing server.

My point isn't that LADDIS is a bad benchmark! To test a server to it's breaking point an I/O intense load is a good thing. And a few real-world loads are as hard as LADDIS.

But don't assume that LADDIS results will limit the performance that you see in a particular real-world environment.

Dave