Cores vs Workers

[kriesel]

Originally Posted by irowiki

Ah, so if you're in it for the long haul, configuring a system to have 1 worker per real core is the best, it'll take "longer" per job but you'll churn out more jobs in the same amount of time, and get the same amount of ghz days more or less?

I have a lot of older dual core systems with no hyper threading. Then a smattering of newer i5/i3's which randomly have or do not have hyperthreading.

One worker per core is not necessarily the optimal for aggregate throughput long term. Benchmark each system. Optimal cores/worker is a function of system details and of fft length for primality testing. I've seen it range anywhere from one worker per core to one worker per multi-core socket, on the same system, in the same benchmark run, as a function of fft length.

One core per worker maximizes latency per primality test assignment, so on a slow system the first-time test may be a double-check by the time it finishes.

[kriesel]

Originally Posted by Prime95

Prime95 always chooses one worker (actually i think it's one worker per four cores). The reason it does this is so that new users will see results "quickly" -- hopefully finishing a work unit before quitting the project.

Some machines are faster with one worker using multiple cores, some are faster with one worker per core. Benchmark to find out.

If it's one per four cores, that's not so good for systems such as dual hex core; one worker would inevitably straddle sockets and take a performance hit.

If prime95 remains at low priority relative to other system activity, we users need to do our part by not using the system for other things during benchmarking, or the fluctuating load will distort the benchmark's results and mislead us as to which core/worker choice produces more prime95 throughput, assuming it's going by wall clock rather than cpu clocks used in the application.

[irowiki]

Originally Posted by Mark Rose

In my experience, on machines before Skylake, one worker per core is almost always better. With Skylake, 1 worker is often better, depending on FFT size (up to 4096k on my systems). The L1/2/3 cache sizes and memory bandwidth also impact which is better.

I usually run 1 worker regardless because I like seeing results faster. The latest Prime95/mprime will do periodic benchmarks to find the most optimal settings, once you've configured the number of workers/core.

Ah, I would say the majority of my machines are pre-Skylake. So i will leave them alone and just play around with a few machines I have easy access to.

Thank you!