2 x Xeon X5690 system. Questions about optimal efficiency.

[armeg]

So I have a few questions about my 2 x Xeon X5690 system, and I want to confirm my understanding/correct any misunderstandings I have.

At the moment, the system is doing PRP (153) work only. On my assignments page, core 1 is doing M(332290513) and core 2 is doing M(332418467).

At the moment I am only running 2 workers, priority 1, and CPU cores to use set to 6 for each. Based on my reading this is not an optimal set-up. It seems that generally, the optimal situation is 1 worker per physical core.

According to lscpu, I have 2 threads per core, 6 cores per socket. So, the optimal solution here would be to run 12 workers (according to my reading/understanding). When I go to attempt to set it to 12 in mprime under Test/Worker threads. It automatically chooses 4 CPU cores as the value to use. This value is impossible since I only have 12 cores in the first place, and a total of 24 threads. If I ignore this and attempt to get to the end I get an error saying:

You have allocated more cores than are available.  This is likely to
GREATLY REDUCE performance.  Do you want to correct this? (Y):

This makes sense considering it's assigning 4 CPU cores without asking me. My question here is a little open-ended I guess. Why is mprime not prompting me to enter CPU cores to use? If it were prompting me, wouldn't I want to set cores to use to 1?

I appreciate any help you guys have for a newbie.

[VBCurtis]

On a 100-million-digit test such as the ones you selected, one worker per socket (rather than per core) is optimal. The idea is that more of the data for the one job will reside in CPU cache, somewhat shared among individual threads on a socket testing the one number.

Even if one worker per core was optimal, the time to complete that first batch of 12 tests would be rather difficult to stomach; even if 6 cores per worker lost a few percent in efficiency, it's worth it to complete a single test faster.

[Prime95]

Originally Posted by armeg

Why is mprime not prompting me to enter CPU cores to use? If it were prompting me, wouldn't I want to set cores to use to 1?

VBCurtis hit the nail on the head. Even if one core per worker is optimal throughput, waiting for 12 single-core tests to complete will be pain-staking. In fact, prime95/mprime won't let you choose less than 4 cores when testing 100-million digit numbers. You can do what you want, but only by manually editing the local.txt file.

[LaurV]

Run a benchmark, and decide after.

Post the results here so we learn something...

[ewmayer]

Originally Posted by Prime95

VBCurtis hit the nail on the head. Even if one core per worker is optimal throughput, waiting for 12 single-core tests to complete will be pain-staking. In fact, prime95/mprime won't let you choose less than 4 cores when testing 100-million digit numbers. You can do what you want, but only by manually editing the local.txt file.

My own rule of thumb for this sort of thing: Use the maximum numbers of threads per job which gives you at least 90% of the maximum-total-throughput configuration.

[aurashift]

Originally Posted by ewmayer

My own rule of thumb for this sort of thing: Use the maximum numbers of threads per job which gives you at least 90% of the maximum-total-throughput configuration.

For 100M exponents, depending on ram and other stuff of course, there's usually a point around the 12 core mark where leaving one core free to handle interrupts from day to day stuff from the OS is beneficial. The curve from throwing more and more cores becomes less beneficial, but then again someone around here (not me) thought more cores but less clock speed would be a good idea around the 18+ core mark, so YMMV. (This appears to have an impact on your memory controller performance) We're going back to high clock speeds but I don't have my hands on those goodies yet.



Source: I used to do 100M exponents exclusively until I got bored.