5% performance gain running NFSNET + PRIME on a P 4 HT

[TauCeti]

Some days ago i joined http://www.nfsnet.org (those folks below in the forums here) because i think they really need some more computing power *g*

Playing with my P4 2.6@ 2.83 System i made the following observations:

The NFSNET client performance is heavily (more then prime95) bound by the main-memory bandwith and does not use SSE2.

So i tried to run prime95 23.4.1 and the NFSNET client simultaniously on my System (WIN XP Pro SP1) with hyperthreading enabled. Prime95 was doing LL-testing on 19906457.

Test 1: Prime and NFSNET on low priority, HT enabled
Prime95: 0.065/iter (0.033 running alone)
NFSNET: 6200 lines/day (11200 running alone)

Test 2: Prime at low priority, NFSNET on 'normal', HT enabled
Prime95: 0.064/iter (0.033 running alone)
NFSNET: 6350 lines/day (11200 running alone)

To check the effect of HT, i disabled it

Test 3: Prime and NFSNET on low priority, HT disabled
Prime95: 0.065/iter (0.033 running alone)
NFSNET: 5900 lines/day (11200 running alone)

*Result*: Running prime and NFSNET simultaniously with default priorities on my HT-P4 system shows prime95 running at 50-52% speed and NFSNET at about 60% speed. That is an improvement of at least 5% for the overall efficiency (compared with two systems running only one client each).

Even without HT you get about 2-3% increase in overall throughput.

I cannot explain why raising the priority of NFSNET gains performance without affecting prime95. Perhaps it's a measurement error. The benchmarking with NFSNET is not that easy like it is with prime95 but the total increase in overall throughput is imho far above the measurement errors.

Cheers!

[ColdFury]

I suspect because Prime95 is fpu-intensive, while the number field sieve is mainly integer operations. (Except maybe for the sieve?)