Which Work Types are Multithreaded

[tului]

I've noticed it seems P-1 is, but what others would benefit from say having 1 main CPU assigned and 3 others for threads. Regardless of real cores or SMT or HT style debate. Just want to know what all work types can and do in the current build use multiple threads for improvement.

[legendarymudkip]

LL, P-1 and ECM are all multithreaded, past a certain point. For P-1 and ECM the cutoff point is around an exponent of 200000, as far as I know. With LL I think the cutoff is very low. I don't know about PRP tests. Trial factoring isn't multithreaded, but GPUs are much better at it, even if it were multithreaded in prime95.

[ATH]

Doing ECM on M1277 seems to receive no benefits from multithreading, adding 4 threads it still only uses same cpu percentage as 1 thread.

I think this is because of the tiny FFT length? Multithreading only benefits higher FFT lengths?

[kladner]

Originally Posted by tului

I've noticed it seems P-1 is, but what others would benefit from say having 1 main CPU assigned and 3 others for threads. Regardless of real cores or SMT or HT style debate. Just want to know what all work types can and do in the current build use multiple threads for improvement.

On an FX8350, when I was running P-1 with two threads each, four Workers, 8 "cores" (integer units which shared an FPU), there was an interesting pattern. Stage One would pretty much keep both "cores" of the set busy, as would ECM. During Stage Two (28 GiB RAM available), the second core would drop to something like ²/₃ of full usage, once the allocated memory for the Worker was fully engaged.

The first screen shot shows a single Stage 2 Worker on the 7th and 8th cores. The other shows the transition from Stage 2 Init for cores 5 and 6.

[ATH]

Hmm this is wierd, doing 4 x ECM on M1277 now it uses 4 times as much cpu as before, but each test takes 60% longer than running 1 ECM with 4 threads, even if the cpu usage was low.

Before it was using 12-13% cpu which is what 1 thread normally uses (4 physical cores, 8 threads with HT which I cannot disable). Now 4 tests is using 50% cpu which is the full cpu when HT is enabled.

But since each test is taking 60% longer now it must have gotten some benefits from multithreading before even though it could not be seen in the cpu usage.

[Prime95]

Originally Posted by ATH

Doing ECM on M1277 seems to receive no benefits from multithreading...

I think this is because of the tiny FFT length? Multithreading only benefits higher FFT lengths?

Yes, only larger FFTs will benefit from multithreading.

[VBCurtis]

Originally Posted by ATH

Hmm this is wierd, doing 4 x ECM on M1277 now it uses 4 times as much cpu as before, but each test takes 60% longer than running 1 ECM with 4 threads, even if the cpu usage was low.

Before it was using 12-13% cpu which is what 1 thread normally uses (4 physical cores, 8 threads with HT which I cannot disable). Now 4 tests is using 50% cpu which is the full cpu when HT is enabled.

But since each test is taking 60% longer now it must have gotten some benefits from multithreading before even though it could not be seen in the cpu usage.

Or HT is decreasing your efficiency... Perhaps you have other tasks going on some cores, or P95 is not properly assigning cores to avoid HT-contention.