AMR Sim on Threadripper (Linux)


#1

Hey, just wanted to make sure that everything works as expected, I’ve encountered some programs that have trouble with an extensive amount of cpu threads.

I’ve set max usage for both Sim and Glonet.

Started Gearing Strategy. The Sim client is frequently jumping between 5-26 used cores(of 64 if including threading) at 3.3GHz and a 1% err rate(everything else is default value). Is gearing strat taking 3 hours for 1359 iterations a normal timespan?

Best regards, Alex


#2

How long it takes depends on the spec, number of cores, and how fast the cores are.

On my Core i7-6850K (6 cores, 12 logical cores), I can run a gearing strategy in under 10 minutes for some DPS specs. For some healer specs like restoration druid, it can take up to an hour or more.

If you have a lot of reasonably fast cores, I would recommend running multiple clients that use less threads per client – due to how the simulator is written, it will distribute the work better. (On my list to handle this kind of thing automatically, but not many people have more than about 8 threads, so keeps getting bumped down the list.) Try running multiple 8-thread clients to use all 64 threads more consistently, see how it goes.

You might find that a different number of clients works best for your CPU – different CPUs have different performance profiles. You would want to run at least the number of physical cores that you have, e.g. if you have 32 physical cores, you’d want to run at least 4 8-thread clients. You would probably see a small increase by going a bit higher due to the gains modern CPUs can get via hyperthreading (or whatever AMD calls the equivalent behavior). Play with the number of clients until you reach the sweet spot for your machine.

For reference, I have an older 8-core Ryzen, and I can hit near-100% CPU usage with 3 8-thread clients these days. Usually 2 is enough, but for various development reasons I run 3.


#3

Thanks, it really helped! Started 3 ubuntu docker containers with the amr client running custom on 8 cores. Uses 28-34 cores in total. Even though the amount of cpu used didn’t change equally, the gearing simulation re-run took only 10m instead of ±3h.

Confirming that maxing on a heavily threaded machine isn’t best practice at the moment.