Anyway, this is a W2100z running Solaris 10 as a client vs a (busy) V880 running Solaris 10 as the NFS server:

time dd if=humanest1.nsq of=/dev/null bs=1048576
404+1 records in
404+1 records out
0.00u 1.52s 0:04.32 35.1%

So that’s about 90 megabytes/s in a single stream over NFS.

You comment about single stream performance is a valid one though. My felling is that single-stream performance has improved in Solaris 10. That’s not my primary focus though - I’m mainly interested in overall NFS server throughput, and it’s clear that NFS is a bit slower than a pure stream over TCP.

I tried a pair of V240s (not quite identical unfortunately) running Solaris 9 and 10 as clients, and it was clear that Solaris 10 makes a difference there. I was getting 5.68s versus 8.91s, with Solaris 10 obviously quicker.

I used ttcp to test the throughput on the Solaris 9 machines (the 280R and E2900). I did some quick-n-dirty NFS testing, but never really paid much attention to the results since they’re so variable. Like I just ran this quick test. Reading a 2GB file on an idle E2900 running Solaris 9(file was created on another host, so it’s not cached on the 2900):

# time dd if=testfile of=/dev/null bs=1048576
2000+0 records in
2000+0 records out

real 0m48.014s
user 0m0.010s
sys 0m12.060s

Roughly 42MB/sec. The filer reports about the same (45000 KB/sec) during the test, and is otherwise relatively idle.

By comparison, from a v20z running Solaris 10 hits about 53MB/sec:

# time dd if=testfile of=/dev/null bs=1048576
2000+0 records in
2000+0 records out

real 0m37.913s
user 0m0.029s
sys 0m8.050s

And a v20z running a Linux 2.4 (!) kernel musters 50MB/sec:

$ time dd if=testfile of=/dev/null bs=1048576
2000+0 records in
2000+0 records out

real 0m40.360s
user 0m0.000s
sys 0m8.540s

Of course, I’ve often seen single Linux hosts push over 95MB/sec to the filer. So right now something is either slowing down my filer or my network. And of course I just did another write to the filer from the 2900 and moved about 65MB/sec. Ah, the joys of testing on a live network.

Anyhow, what I’m interested in in particular is ’single stream’ throughput. One thing is for sure, it’s got me motivated to do some more scientific testing with a dedicated GigE switch.

Certainly our Solaris 10 (V440, W2100z) machines with gigabit can be clocked at something like 990Mbit/s using a single rcp. (Ouch!) I just checked a V65x running Solaris 9 and it managed to give me 885Mbit/s sustained.