Is TF from 2^64 to 2^65 using SSE2?

TauCeti · 2003-10-16, 13:07

I know that whatsnew.txt states for Version 22.7 : "Trial factoring above 2^64 on a P4 now uses the SSE2 instructions."

I just want to confirm here, that this is really the case, because i get some strange results using Hyperthreading simultanously with one LL-instance and one TF-instance.
Actually, between 2^64 and 2^65 i get a net-throughput increase of roughly 44% using HT compared to 20% between 2^63 and 2^64.

Before i started to test HT, i expected a performance _decrease_, once TF and LL both start using SSE2.

If 2^65 on P4 is really SSE2, then more on this topic from me in another thread in the hardware-section later...

Tau

TauCeti · 2003-10-16, 15:30

... hmmm, my strange results are still there but now i have tested a 2^67 TF with the same result and i am pretty sure SSE2 is used there. More in the hardware section...

Tau

Prime95 · 2003-10-16, 22:05

The FFT is using floating point SSE2 instructions. Trial factoring is using the integer SSE2 instructions. It makes sense that HT would benefit in this scenario.

TauCeti · 2003-10-17, 06:30

Thanks for the explanation. Maybe i really shold try to understand SSE2 in detail

So - assuming for the sake of the argument - that only one HT-machine alone participates in GIMPS, would it not improve overall GIMPS throughput if HT-Users could choose in the client to use a different table for the TF-limits because in:

factoring_cost < chance_of_finding_factor * 2 * primality_test_cost

the factoring cost is reduced by about 40% using HT?

Tau

2003-10-16, 13:07	#1
TauCeti Mar 2003 Braunschweig, Germany E2₁₆ Posts	Is TF from 2^64 to 2^65 using SSE2? I know that whatsnew.txt states for Version 22.7 : "Trial factoring above 2^64 on a P4 now uses the SSE2 instructions." I just want to confirm here, that this is really the case, because i get some strange results using Hyperthreading simultanously with one LL-instance and one TF-instance. Actually, between 2^64 and 2^65 i get a net-throughput increase of roughly 44% using HT compared to 20% between 2^63 and 2^64. Before i started to test HT, i expected a performance _decrease_, once TF and LL both start using SSE2. If 2^65 on P4 is really SSE2, then more on this topic from me in another thread in the hardware-section later... Tau

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Athlon 64 and SSE2	ThomRuley	Hardware	17	2004-05-14 19:26
RSA and SSE2	Cyclamen Persicum	Math	5	2003-11-10 07:41
P4 SSE2 routine bug?	TTn	Lounge	27	2003-07-17 17:14
SSE2 ?	TauCeti	NFSNET Discussion	8	2003-06-30 12:58
The effect of SSE2 in P4s	cmokruhl10	Hardware	8	2003-06-17 11:18

2003-10-16, 15:30	#2
TauCeti Mar 2003 Braunschweig, Germany 2×113 Posts	... hmmm, my strange results are still there but now i have tested a 2^67 TF with the same result and i am pretty sure SSE2 is used there. More in the hardware section... Tau

2003-10-16, 22:05	#3
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 2×4,079 Posts	The FFT is using floating point SSE2 instructions. Trial factoring is using the integer SSE2 instructions. It makes sense that HT would benefit in this scenario.

2003-10-17, 06:30	#4
TauCeti Mar 2003 Braunschweig, Germany 2×113 Posts	Thanks for the explanation. Maybe i really shold try to understand SSE2 in detail So - assuming for the sake of the argument - that only one HT-machine alone participates in GIMPS, would it not improve overall GIMPS throughput if HT-Users could choose in the client to use a different table for the TF-limits because in: factoring_cost < chance_of_finding_factor * 2 * primality_test_cost the factoring cost is reduced by about 40% using HT? Tau