Testing complete
2 sets/client 2x clients on 2 CPUs (HT OFF) under Fedora Core 1 (2.4.22-1.2115.nptlsmp) run:
Total run time: 283030s (78:37:10)
Client 0: full cycle, 233:71; 68371
Client 1: full cycle, 198:31; 64831
Total: 133202 structures
Avg rate: 28.24 structures/min
I just realized that my first run spilling over into the 'easy' gen0 structures could skew the results some, so I have a 'handicapped' number below, where I made my calculations using the same total run time, but did not give credit for any of the gen0 structures.
In the end...it looks like HT gives a boost of about 14.9-19.1% (when used properly). When used improperly, it yields about a 10.5% performance defecit (this assumes an otherwise idle machine, very likely in Linux, highly unlikely in Windows, due to GUI overhead).
Well, I'm off to start a few more clients on some Xeons... ;p
33.64 HT ON, 1x client per virtual CPU
32.45 HT ON, 1x client per virtual CPU (handicapped)
28.24 HT OFF, 1x client per physical CPU
25.28 HT ON, 1x client per physical CPU
__________________
Everything aircooled  (I honestly don't know why you guys let me stay a member...  )
Always folding:
WXP Pro:
KD7R + Palomino 1800+
W2K3:
4x P3 Xeon 500 512KB (4 clients)
Linux boxes:
none  Had to put them to use doing *real* work.
|