I guess that previous Javascript benchmark episode set me into benchmarking mode!  Combine that with a discussion on the Infrant forums about NFS over UDP vs. NFS over TCP and I knew I had to do some testing!  Once the NFS battle was settled, I figured I might as well expand to include some other drives available on the same machine (Sun v250, dual 1.28ghz processors and 8GB of RAM)...  The first test was the time to checkout a decent sized Subversion code repository from a gigabit connected AMD Sempron running at 2ghz over the svn protocol, and the second time was what it took to "rm -rf" the tree. So, here's the results, in standard "time" output format (real, user, sys):

NFS over UDP: Checkout = 7:50, 1:18, 0:40.  Remove = 1:31, 0:00.292, 0:56.4

NFS over TCP: Checkout = 7:09, 1:18, 0.30.  Remove = 0:48, 0:00.220, 0:22.9

Local ZFS (3 drive RAIDZ):  Checkout = 2:22, 1:15, 0:25.  Remove = 0:00.92, 0:00.118, 0:00.779

Local UFS (2 drive RAID1):  Checkout = 3:58, 1:16, 0:20.  Remove = 0:27.9, 0:00.14, 0:01.289

I was surprised to see TCP beating UDP, but apparently the fact that TCP windows packets MORE than makes up for the processing overhead on high speed links!  Another shocker was to see the performance of ZFS vs. UFS!  I had expected RAIDZ, with it's distributed parity, to be slower than UFS on a hardware mirror!  Guess that's why we actually do the tests!  :-)