Hyperthreading and CPU utilization.

[cheesehead]

Originally Posted by MS63

if one runs two instances of Prime95, do they each require their own User/Computer ID?

No, not required. PrimeNet (and George) will happily accept results with a single user/computer ID that were produced on multiple CPUs or systems -- and lump them all in the same total of factoring and LL time for ranking purposes.

But if you want PrimeNet to keep separate totals for different instances, then you'll need to use separate IDs.

[MS63]

I seem to remember that a problem with UNassigning occurs if one allows PrimeNet to make assignments and both instances of P95 have the same User and Computer ID's.

For example, P95a has 1234567 assigned and P95b has 123456789. When, say, P95a contacts PrimeNet, the candidate 123456789 will be UNassigned. Then, when P95b contacts PrimeNet the 'not assigned' error will be produced.

Or vice versa.

If someone says I'm wrong then, unless I can be bothered to test it again, I'm wrong. Even then I could be wrong.

[patrik]

Well, I'd say you are wrong, since this is how I'm running one of my computers without any problems.

Regarding the iteration times, I think it is different for different processors, and maybe also for different versions of Prime95 (whether or not running two instances is better than one).

[cheesehead]

Originally Posted by MS63

I seem to remember that a problem with UNassigning occurs if one allows PrimeNet to make assignments and both instances of P95 have the same User and Computer ID's.

Oh, I think you're right.

I was thinking of situations where all my PrimeNet assignments were executed on a single system. I reported results from non-PrimeNet-assigned work on other systems using the same user & computer ID.

[MS63]

It's one apiece. (Sort of.)

Any more views on this?

[MS63]

Originally Posted by MS63

From VERY unreliable memory my per iteration time INCREASED when I ran two instances of Prime95, not decreased as you have experienced.

I will experiment with this again after Christmas: ...

Well, as promised, here are my findings:

I am running an HT enabled Pentium 4E 3.0 GHz Prescott 90nm processor with 16KB L1 cache and 1024KB L2 cache. There are 2GB of RAM installed.

The two exponents involved in my testing were 15002213 (Pa) and 16147193 (Pb). I cannot recall the FFT sizes.

Running alone, with or without affinity set, Pa had a per iteration time of 0.028 sec.

Running alone, with affinity set to both none and 1, Pb had a per iteration time of 0.032 sec, and a per iteration time of 0.033 sec with affinity set to 0.

With both Pa and Pb running the best mean per iteration time of 0.02775 sec was obtained when both Pa and Pb had affinity set to none, or both had affinity set to different CPU's, or Pa was set to affinity 1 and Pb set to affinity none. The individual per iteration times for Pa and Pb varied, but as one increased the other decreased by the same amount.

Other affinity combinations resulted in an increased mean per iteration time, up to 0.02875 sec for affinity Pa set to 0 and affinity Pb set to none.

So, in summary, for my computer the best settings for the two exponents given above are to have two instances of Prime95 running with Pa affinity set to 1 and Pb affinity set to 0. (These affinity settings being for my own reasons.)

[drew]

Originally Posted by MS63

Well, as promised, here are my findings:

I am running an HT enabled Pentium 4E 3.0 GHz Prescott 90nm processor with 16KB L1 cache and 1024KB L2 cache. There are 2GB of RAM installed.

The two exponents involved in my testing were 15002213 (Pa) and 16147193 (Pb). I cannot recall the FFT sizes.

Running alone, with or without affinity set, Pa had a per iteration time of 0.028 sec.

Running alone, with affinity set to both none and 1, Pb had a per iteration time of 0.032 sec, and a per iteration time of 0.033 sec with affinity set to 0.

With both Pa and Pb running the best mean per iteration time of 0.02775 sec was obtained when both Pa and Pb had affinity set to none, or both had affinity set to different CPU's, or Pa was set to affinity 1 and Pb set to affinity none. The individual per iteration times for Pa and Pb varied, but as one increased the other decreased by the same amount.

Other affinity combinations resulted in an increased mean per iteration time, up to 0.02875 sec for affinity Pa set to 0 and affinity Pb set to none.

So, in summary, for my computer the best settings for the two exponents given above are to have two instances of Prime95 running with Pa affinity set to 1 and Pb affinity set to 0. (These affinity settings being for my own reasons.)

To revive an old thread...I just realized the flaw in the strategy of testing slightly faster with 2 instances of Prime95. My hard drive just crashed, and now I've lost twice as much work as I would have otherwise. From the few percent performance advantage of running two instances of Prime95, It'd be quite awile before I make up for the additional work lost.

From now on, I'm going to run one test at a time...finish them faster and less will be at risk. Lesson learned.

Drew

[Cruelty]

To backup or not to backup - this is a question

[drew]

Originally Posted by Cruelty

To backup or not to backup - this is a question

Yes, in hindsight that would have helped, but except for Prime95 I don't keep anything on that PC worth backing up.

What I could do is write a script that will transfer the work files between my two networked PCs once in awhile.