running mprime on 4 cores on 2 assignments?

[tha]

Is there a way to force mprime so core 1&2 run on one assignment and core 3&4 on a second assignment?

[kladner]

On an FX-8350 (8 cores), I set Number of Workers to 4, and CPUs to use to 2. When I tried this, assigning work to the odd-numbered cores resulted in the even numbers being assigned as helpers.

Intel may be different. Others will know about that.

(And oops. My experience is with P95. It's been a while since I ran Mprime, so I'm not sure of the procedures.)

[petrw1]

You can use AFFINITYSCRAMBLE2 to pick your own cores.

Read all about it in undoc.txt

At least you can in Prime95...not sure if mprime is different

[Mark Rose]

Originally Posted by petrw1

You can use AFFINITYSCRAMBLE2 to pick your own cores.

Read all about it in undoc.txt

At least you can in Prime95...not sure if mprime is different

You can. The only difference is that Windows numbers the CPU cores differently from what I remember.

[TObject]

That is right.
Cores on Windows: Physical, Virtual, Physical, Virtual, Physical, Virtual, Physical, Virtual
Cores on Linux: Physical, Physical, Physical, Physical, Virtual, Virtual, Virtual, Virtual

[tha]

These are my timings for a 4 core non hyperthreaded cpu:

Timings for 1792K FFT length (1 cpu, 1 worker): 33.98 ms. Throughput: 29.43 iter/sec.
Timings for 1792K FFT length (2 cpus, 2 workers): 41.41, 41.39 ms. Throughput: 48.31 iter/sec.
Timings for 1792K FFT length (3 cpus, 3 workers): 50.30, 49.11, 39.11 ms. Throughput: 65.81 iter/sec.
Timings for 1792K FFT length (4 cpus, 4 workers): 59.27, 60.37, 60.64, 59.51 ms. Throughput: 66.73 iter/sec.

There is little use for a fourth core as one can see.

I tried fiddling with all the info from undoc.txt but failed to get two cpus to work on a single assignment.

What I would like to try is a 4 cpus, 2 workers option, but I don't know if that can be done by mprime or prime95.

[LaurV]

Originally Posted by tha

These are my timings for a 4 core non hyperthreaded cpu:

Timings for 1792K FFT length (1 cpu, 1 worker): 33.98 ms. Throughput: 29.43 iter/sec.
Timings for 1792K FFT length (2 cpus, 2 workers): 41.41, 41.39 ms. Throughput: 48.31 iter/sec.
Timings for 1792K FFT length (3 cpus, 3 workers): 50.30, 49.11, 39.11 ms. Throughput: 65.81 iter/sec.
Timings for 1792K FFT length (4 cpus, 4 workers): 59.27, 60.37, 60.64, 59.51 ms. Throughput: 66.73 iter/sec.

There is little use for a fourth core as one can see.

I tried fiddling with all the info from undoc.txt but failed to get two cpus to work on a single assignment.

What I would like to try is a 4 cpus, 2 workers option, but I don't know if that can be done by mprime or prime95.

Yes, it can. Your 4th worker is memory bounded. You can go to Test menu, select 2 workers, in the "worker number" select "all", and in the "cpus to use" select 2. Then you'll have 2 workers, each using 2 threads. You can also do other tricks, like a single worker, using 4 "cpu"s (=threads, cores).

[Dubslow]

Originally Posted by tha

I tried fiddling with all the info from undoc.txt but failed to get two cpus to work on a single assignment.

What I would like to try is a 4 cpus, 2 workers option, but I don't know if that can be done by mprime or prime95.

The Prime95 GUI can do it, or with Mprime you can do it manually by setting the WorkerThreads and ThreadsPerTest variables in local.txt. (The Prime95 GUI is merely a frontend -- its effects is exactly to set these variables).

To get two workers going with two cores each, set it thusly:

WorkerThreads=2
ThreadsPerTest=2

At the bottom of the file you can set the individual affinities:

[Worker #1]
Affinity=0

[Worker #2]
Affinity=2

That sets one worker on core 0, the other one on core 2. You may additionally use the AffinityScramble2 option to define where the helper threads run:

AffinityScramble2=0123

This should mean (you can verify by searching for this option in undoc.txt) that the worker running on core 0 will have it's helper thread running on core 1, and the worker on core 2 will have its helper thread on core 3. (For example, on a GNU-Linux machine with 4 physical cores and 8 threads, i.e. cores 0-3 are physical and 4-7 are virtual, you can pair each virtual core to its physical partner with the setting AffinityScramble2=04152637.)

[tha]

This worked, thanks for the help. I tried it on a Prime95 version first and than managed to do the same using the -m option with mprime.

The result was interesting. Using 4 cores with 4 workers it does 4 assignments in exactly 20 days. Using 4 cores with 2 workers and 2 cores per worker it does 2 assignments in exactly 10 days.

So the total throughput is the same.

[Dubslow]

Originally Posted by tha

The result was interesting. Using 4 cores with 4 workers it does 4 assignments in exactly 20 days. Using 4 cores with 2 workers and 2 cores per worker it does 2 assignments in exactly 10 days.

So the total throughput is the same.

I would interpret that as "the lost production due to choked memory bandwidth with 4 cores is about the same as the lost production due to multithreading overhead".