New PC can handle two instances of Prime 95

[Bundu]

I think this topic has been talked about before but I was drunk when I saw it!

I've just bought a new PC, P4 - 3.2Ghz with Hyperthreading - 512Mb RAM - 1Mb L2 cache. Whenever I run a test (factoring or 10,000,000+ I find that only 50% of my CPU is used by Prime95)

Do you guys think that I could successfully run two instances of Prime95 (one doing LL at priority2 and one doing factoring at priority 1)?

I'm just trying to find out if anyone has similiar specs and has tried or if I should just "go it alone" and test it out

Thanks, Mark

[S80780]

Hi, Mark!

Indeed, these 50% displayed are 100%, as the virtual Processor is regarded as a 2nd idle Hardware-Processor, dispite the hardware behind it doesn't exsist.
There were discussions on getting a higher over-all-throughput by running two instances of Prime95, one doing LL-Tests and one doing TF. Nonetheless, this needs some hand-tuning and as I don't have a P4 at hand, I can't give you instructions on how to do it.

Benjamin

[scottsaxman]

I run two instances of the NSFNET client on my 3.2 P4 w/ HT. It doesn't double the output, but perhaps adds another 70% or so.

[geoff]

There is a definite throughput advantage in running trial factoring simultaneously with LL testing on a HT enabled P4. If you do this I would recommend running two instances of the torture test as a check. I had to reduce the overclock on my 2.4 P4 by about 5% to get it stable with hyperthreading enabled.

[ET_]

Originally Posted by geoff

There is a definite throughput advantage in running trial factoring simultaneously with LL testing on a HT enabled P4. If you do this I would recommend running two instances of the torture test as a check. I had to reduce the overclock on my 2.4 P4 by about 5% to get it stable with hyperthreading enabled.

If you run factoring over 65 bit on a Pentium IV, SSE2 optimizations will lead you to an hardware contention with LL-tests, and this definitely throws down performance.

Luigi

[Bundu]

Originally Posted by scottsaxman

I run two instances of the NSFNET client on my 3.2 P4 w/ HT. It doesn't double the output, but perhaps adds another 70% or so.

NSFNET??

[Bundu]

Originally Posted by ET_

If you run factoring over 65 bit on a Pentium IV, SSE2 optimizations will lead you to an hardware contention with LL-tests, and this definitely throws down performance.

Luigi

I'm going to take that all that as "erm... not a good idea Mark"??

[PrimeCruncher]

Originally Posted by Bundu

I'm going to take that all that as "erm... not a good idea Mark"??

People have benchmarked HT CPUs before. They got better overall output, but each instance of Prime95 by itself was slower. (E.g. instance one = 70% of normal and instance two = 70% of normal so 140% throughput) P4s stink, badly, at factoring below 64 bits, but I guess if you really wanted to contribute, you could run TF or LL as instance one and do LMH as instance two.

[geoff]

Originally Posted by ET_

If you run factoring over 65 bit on a Pentium IV, SSE2 optimizations will lead you to an hardware contention with LL-tests, and this definitely throws down performance.

I have tested it with double checking LL (768K FFT) and TF at above 65 bit and throughput is about 10% better than running separate instances. I haven't tried larger FFT sizes though.

I think the LL tests use floating point SSE2 whereas factoring uses integer SSE2.

edit: (This is on linux 2.6, mprime 23.5, P4 2.4C)

[jocelynl]

Be carefull when using 2 instances. This will increase the temperature of your cpu and could generate false results. Make sure you run many FFT test first.

Joss