Multiple Core for 1 assignment

[saeres]

Is there a program that allows one to use multiple cores (in cpu) to complete an assignment. I read that someone managed to verify a mersenne prime using 32 server cores and wanted to take advantage of such utility.

[Uncwilly]

The current version of Prime95 can use many cores on a single assignment.

[saeres]

I'm running 3 assignments on 1 quad core and they might reach 80% of total cpu power. I'm looking for 1 assignment to be computed by near 100% of cpu (1 assignment might us 30% max)

[saeres]

Nevermind. For some reason prime95 showed that I had only dual core and now it's registering all 4 cores and is allowing me to address the assignment to all four cores.

[Uncwilly]

Just so that you know:

You will get more assignments done for a given time (assuming it is long enough to be significant) by having each core work on its own assignment. Having 4 cores work on a single assignment will not get 4x the speed of a single core. It will be closer to 2.7x faster. There are others that have the data to demonstrate this. Run the tests. Use a single core and note the iteration times, then 2 cores, then 3, then 4.

[billvau]

Uncwilly - just saw your post on assigning cores. I'm a newbie that just started 2 days ago. I thought that in setting up the program I had to tell it the number of CPUs to use but now I can't find that. Can you help me by pointing me to configuration documentation? Or, do you know a "best" configuration?

Thanks much,
Bill

[kladner]

To set workers, go to (menu) Test>Worker windows. Welcome to the project!

[billvau]

Thanks, kladner!

[cuBerBruce]

Originally Posted by Uncwilly

Just so that you know:

You will get more assignments done for a given time (assuming it is long enough to be significant) by having each core work on its own assignment. Having 4 cores work on a single assignment will not get 4x the speed of a single core. It will be closer to 2.7x faster. There are others that have the data to demonstrate this. Run the tests. Use a single core and note the iteration times, then 2 cores, then 3, then 4.

Beware of what Uncwilly is saying. He is comparing 1 worker with 4 cores versus 1 worker with 1 core, rather than comparing 4 workers with 1 core each versus 1 worker with 4 cores.

On my Haswell system, I get much less than 4x throughput when running 4 workers on 1 core each versus running 1 worker on a single core, just like Uncwilly's example. However, I get basically the same throughput running 1 worker with 4 cores as I do running 4 workers using 1 core each. My throughput seems to be only dependent on the number of physical cores being used, and essentially no significant effect due to cores per worker (as long as only 1 thread per physical core is being used).

You also need to make sure each thread is running on a separate physical core. I have given up on relying "smart assignment" to do the right thing, and simply edit my local.txt to make sure affinity is set up properly.

Also, do not use CPU usage percentage as a measure of how much work is actually getting done.

[chalsall]

Originally Posted by cuBerBruce

You also need to make sure each thread is running on a separate physical core. I have given up on relying "smart assignment" to do the right thing, and simply edit my local.txt to make sure affinity is set up properly.

I completely agree with you. After a great deal of experimentation, I found that running 4 tests on 4 real CPUs generated about the same net throughput as running 1 test on 4 real CPUs.

Can anyone say memory/cache bottleneck?

However, I find it a bit disappointing that Prime95 / mprime can't seem to figure out the optimal affinity on its own. When you have threads jumping around all the available CPUs (including the virtual ones) your net throughput is significantly affected.

[LaurV]

Originally Posted by cuBerBruce

I get basically the same throughput running 1 worker with 4 cores as I do running 4 workers using 1 core each

This only shows that your system does not have a memory bandwidth bottleneck, for the range of the exponents (FFT size) you are doing, or if it has one, it manifests itself in both cases. For the former, increasing the exponent will start showing differences in favor of 1 worker 4 cores, and for the last, decreasing the exponent will start showing differences in favor of 4 workers one core each. For all the systems I ever played with, the higher output flow I could get was (almost) always n workers with 1 core each, where n is the number of physical cores. With very seldom exceptions, and strange systems, sometime you can get few percents by using HT (doubling the number of workers or using 2 cores for each worker), with the cost that your system is much hotter, less responsive, less stable, and the CPU consumes almost double, etc. But generally, the best is when you go n workers, 1 core each, this only if your system has enough memory bandwidth. Again, this depends of the exponent, cache size, etc.
[edit: for the records, I use water cooling mostly, and I do overclock]