How to get mprime to run self bench? - Page 3

timbit · 2022-06-30, 00:59

Here's 14 workers, 1 core each:

Prime95 64-bit version 30.7, RdtscTiming=1
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=4 (14 cores, 14 workers): 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22 ms. Throughput: 63688.19 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=2 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21 ms. Throughput: 66648.24 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=1 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.22, 0.21, 0.21 ms. Throughput: 65654.27 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=4 (14 cores, 14 workers): 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27 ms. Throughput: 52101.51 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=2 (14 cores, 14 workers): 0.24, 0.23, 0.24, 0.23, 0.24, 0.23, 0.23, 0.24, 0.23, 0.24, 0.23, 0.24, 0.24, 0.24 ms. Throughput: 59580.44 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=1 (14 cores, 14 workers): 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.22, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23 ms. Throughput: 61841.43 iter/sec.

It appears mprime keeps selecting pass1=768, pass2=64, clm=2 for some reason. It's the 2nd slowest one.

timbit · 2022-06-30, 01:31

And here's 30.8b15:

Prime95 64-bit version 30.8, RdtscTiming=1
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=4 (14 cores, 14 workers): 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22 ms. Throughput: 63942.02 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=2 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21 ms. Throughput: 66820.75 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=1 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21 ms. Throughput: 66012.58 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=4 (14 cores, 14 workers): 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27 ms. Throughput: 52146.49 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=2 (14 cores, 14 workers): 0.23, 0.23, 0.24, 0.23, 0.24, 0.23, 0.24, 0.23, 0.23, 0.23, 0.24, 0.24, 0.24, 0.24 ms. Throughput: 59571.93 iter/sec.
FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=1 (14 cores, 14 workers): 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.22, 0.23, 0.23, 0.23, 0.23, 0.23 ms. Throughput: 62050.87 iter/sec.

Also attached.

Prime95 · 2022-06-30, 03:28

Quote:

Originally Posted by timbit

Here's 14 workers, 1 core each:

and what happens now if you run mprime with 14 workers?

timbit · 2022-06-30, 16:45

Quote:

Originally Posted by Prime95

and what happens now if you run mprime with 14 workers?

Looks like all 14 workers are using Pass1=256, Pass2=192, clm=2. Please see attached prime_log_307.txt.

Ok, according to the results.bench.txt for 14 workers, 1 core each, this is indeed the fastest implementation.

But for other configurations, which may not be ideal (running 14 workers 1 core each, stage 2 ECM would take ~8GB each, and my system only has 32 GB total), is there anything else I can do?

kriesel · 2022-06-30, 18:22

Try 7 workers x 2 cores each.
If that alone is not satisfactory, try limiting memory used per worker. In undoc.txt:

Code:

The Memory=n setting in local.txt refers to the total amount of memory the
program can use.  You can also put this in the [Worker #n] section to place
a maximum amount of memory that one particular worker can use.

You can set MaxHighMemWorkers=n in local.txt.  This tells the program how
wany workers are allowed to use lots of memory.  This occurs doing stage 2
of P-1, P+1, or ECM on medium-to-large numbers.  Default is available memory / 1GB.

You can set a threshold for what is considered lots of memory in MaxHighMemWorkers
calculations.  In local.txt, set:
    HighMemThreshold=n        (default is 50)
The value n is in MB.

If you were to limit each worker to 6GB, that would not much reduce the efficiency of stage 2, and allow 4 stage2s running simultaneously. A fifth would get the remaining 4 of your 28, and be less than certain to occur. Over time some workers waiting for sufficient ram to run stage 2 would desynchronize the workers so that not all run stage 1 at the same time, but the set of workers at any given moment run a mix of stage 1 and 2. When one completes stage 2 its memory goes back into the pool to give out to another worker later.

timbit · 2022-07-19, 19:18

I just noticed now that on the mersenne.org "Download software" page, it links to 30.8 b15.
On the main page, the latest announcement lists version 30.7 released, so I never bothered to check for latest stable software release. I guess I need to check the downloads page more often.
I'm still experimenting with the timings, but I have taken the advice and for small FFTs only using 1 or 2 threads max per worker. Throughput has increased, thanks.

2022-06-30, 18:22	#27
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 2³×7²×17 Posts	Try 7 workers x 2 cores each. If that alone is not satisfactory, try limiting memory used per worker. In undoc.txt: Code: The Memory=n setting in local.txt refers to the total amount of memory the program can use. You can also put this in the [Worker #n] section to place a maximum amount of memory that one particular worker can use. You can set MaxHighMemWorkers=n in local.txt. This tells the program how wany workers are allowed to use lots of memory. This occurs doing stage 2 of P-1, P+1, or ECM on medium-to-large numbers. Default is available memory / 1GB. You can set a threshold for what is considered lots of memory in MaxHighMemWorkers calculations. In local.txt, set: HighMemThreshold=n (default is 50) The value n is in MB. If you were to limit each worker to 6GB, that would not much reduce the efficiency of stage 2, and allow 4 stage2s running simultaneously. A fifth would get the remaining 4 of your 28, and be less than certain to occur. Over time some workers waiting for sufficient ram to run stage 2 would desynchronize the workers so that not all run stage 1 at the same time, but the set of workers at any given moment run a mix of stage 1 and 2. When one completes stage 2 its memory goes back into the pool to give out to another worker later. Last fiddled with by kriesel on 2022-06-30 at 18:27

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Running fstrim on SSD while mprime is running might cause errors in mprime	AwesomeMachine	Software	4	2021-10-07 23:49
Radeon VII on a mining-like bench	Viliam Furik	Viliam Furik	17	2021-01-14 08:12
mprime from git	SELROC	Software	2	2018-10-30 10:16
2 x AMD Opteron 2427 @ 2.39 GHz - prime95 bench-	joblack	Hardware	2	2010-03-12 19:38
Problem with mprime (Fixed with mprime -d)	antiroach	Software	2	2004-07-19 04:07

2022-06-30, 00:59	#23
timbit Mar 2009 2²×5 Posts	Here's 14 workers, 1 core each: Prime95 64-bit version 30.7, RdtscTiming=1 FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=4 (14 cores, 14 workers): 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22, 0.22 ms. Throughput: 63688.19 iter/sec. FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=2 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21 ms. Throughput: 66648.24 iter/sec. FFTlen=48K, Type=3, Arch=4, Pass1=256, Pass2=192, clm=1 (14 cores, 14 workers): 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.21, 0.22, 0.21, 0.21 ms. Throughput: 65654.27 iter/sec. FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=4 (14 cores, 14 workers): 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27, 0.27 ms. Throughput: 52101.51 iter/sec. FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=2 (14 cores, 14 workers): 0.24, 0.23, 0.24, 0.23, 0.24, 0.23, 0.23, 0.24, 0.23, 0.24, 0.23, 0.24, 0.24, 0.24 ms. Throughput: 59580.44 iter/sec. FFTlen=48K, Type=3, Arch=4, Pass1=768, Pass2=64, clm=1 (14 cores, 14 workers): 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23, 0.22, 0.23, 0.23, 0.23, 0.23, 0.23, 0.23 ms. Throughput: 61841.43 iter/sec. It appears mprime keeps selecting pass1=768, pass2=64, clm=2 for some reason. It's the 2nd slowest one.

2022-07-19, 19:18	#28
timbit Mar 2009 2²×5 Posts	I just noticed now that on the mersenne.org "Download software" page, it links to 30.8 b15. On the main page, the latest announcement lists version 30.7 released, so I never bothered to check for latest stable software release. I guess I need to check the downloads page more often. I'm still experimenting with the timings, but I have taken the advice and for small FFTs only using 1 or 2 threads max per worker. Throughput has increased, thanks.