View Single Post
Old 2016-09-15, 16:57   #109
airsquirrels
 
airsquirrels's Avatar
 
"David"
Jul 2015
Ohio

11×47 Posts
Default

Quote:
Originally Posted by Madpoo View Post
In Prime95 this shows up VERY clearly if you take a smaller exponent (let's say something in the 5M range) with a small FFT size. If you then run that on a 14-core worker, what happens is that each CPU might only be running at 50% or even less. George explained that was due to the aforementioned inefficiencies in the code... the "master" thread is busy coordinating data from helper threads, or the work chunks aren't being split evenly so some cores are idle while others are still doing things.

It's one of those situations where it actually gets better (more efficient) as the FFT size increases, to the point where at 2M+ FFT sizes, a worker with a lot of cores can manage to keep each CPU at near 100% usually.

So I'm wondering to myself if the greatly larger # of cores involved in KNL is making this effect known again even at the 2M FFT size? Would it scale better at, say, 18M FFT?

One way to see is to throw a test on there in the 332M range and see how it's doing with varying # of cores in the worker, and whether it's able to keep all of the CPUs sufficiently active.

64 physical cores, ~56% utilization
[Work thread Sep 15 12:49] Iteration: 3000 / 332225239 [0.00%], ms/iter: 17.638, ETA: 67d 19:44

32 x 2: ~84% average utilization, similar performance
[Worker #2 Sep 15 12:53] Iteration: 1000 / 332225261 [0.00%], ms/iter: 18.978, ETA: 72d 23:20
[Worker #1 Sep 15 12:53] Iteration: 6000 / 332225239 [0.00%], ms/iter: 18.914, ETA: 72d 17:23

16 x 4: ~96% utilization, performance drops per exponent, overall throughput is up.
[Worker #1 Sep 15 12:56] Iteration: 9000 / 332225239 [0.00%], ms/iter: 32.457, ETA: 124d 19:13
[Worker #3 Sep 15 12:56] Iteration: 2000 / 332225503 [0.00%], ms/iter: 32.663, ETA: 125d 14:15
[Worker #4 Sep 15 12:56] Iteration: 2000 / 332225797 [0.00%], ms/iter: 32.563, ETA: 125d 05:04
[Worker #2 Sep 15 12:57] Iteration: 5000 / 332225261 [0.00%], ms/iter: 32.546, ETA: 125d 03:27
airsquirrels is offline   Reply With Quote