mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Software (https://www.mersenneforum.org/forumdisplay.php?f=10)
-   -   Prime95 30.7 (https://www.mersenneforum.org/showthread.php?t=27180)

Aexoden 2021-10-27 17:22

I'm getting really inconsistent performance on my Threadripper PRO 3955WX (16 cores) running Windows 10 for Workstations. I've only ever used 30.7 builds on this machine, so I don't have any earlier points of comparison (but will certainly test earlier versions if it would matter).

I'm using one worker with 8 threads. Recently, when doing some double checks, it would bounce between either ~2.1ms/iter or ~2.8ms/iter (with no other processes heavily using CPU). Once it was in a particular range, it wouldn't change during a run, but might change modalities after an auto-benchmark or manually stopping/starting. (Though I haven't seen the faster mode since the most recent boot which included a BIOS update and a Windows update.) If other programs were using the CPU, those would slow incrementally (2.1 to maybe as low as 2.5 and 2.8 to maybe 3.2), but there were two clear patterns. Manually assigning affinities in local.txt did help to some extent as well.

Now, having moved on to a larger PRP test in the 110M range, a throughput test on the same settings suggested ~6.6ms/iter, but the test itself was getting ~8.7ms/iter. After restarting the test after removing my Affinity= line, the test now decreased to ~8.0ms/iter, but either I'm misunderstanding the benchmark or there's still something strange going on.

Anyway, if anyone has any ideas or can explain what I'm missing, I'd be all ears. Don't seem to have any similar issues on my 3700X or Threadripper 3960X, but both of those machines are running Linux.

Prime95 2021-10-28 02:12

Build 7 is available. Build 6 does not exist (except for Kruoli).

There is still at least one bug remaining. Running multiple ECM workers can crash with either an out-of-memory event or a restarting with new memory settings event. I've been running for a week under the debugger without any luck. Kruoli has seen one crash in that time.

cyberloner 2021-10-28 10:24

all latest prime95 will fail to run with core cycler .....

gLauss 2021-10-28 12:32

[QUOTE=gLauss;591006]
[C]
[Worker #2 Oct 18 17:03] M6025297 stage 2 is 100.00% complete. Time: 17.558 sec.
[Worker #1 Oct 18 17:03] M6058333 stage 2 is 0.15% complete. Time: 28.509 sec.
[Worker #2 Oct 18 17:03] M6025297 stage 2 is 100.00% complete. Time: 18.477 sec.
[Worker #2 Oct 18 17:03] M6025297 stage 2 is 100.00% complete. Time: 17.475 sec.
[/C][/QUOTE]
can anybody help why this has happened? could not reproduce it so far, see original post for more details...

Prime95 2021-10-28 15:58

[QUOTE=gLauss;591848]can anybody help why this has happened? could not reproduce it so far, see original post for more details...[/QUOTE]

The stage two 100% complete error should be harmless. I'm looking into why it occurs.

kruoli 2021-10-29 10:34

Wavefront P-1, In build 5, when day and night memory settings are different, stage 2 cannot be resumed from the larger memory part:
[CODE][Work thread Oct 29 06:19] M107021149 stage 2 is 45.45% complete. Time: 1957.403 sec.
[Work thread Oct 29 06:52] M107021149 stage 2 is 53.92% complete. Time: 1969.442 sec.
[Work thread Oct 29 07:00] Restarting worker with new memory settings.
[Work thread Oct 29 07:00] Optimal P-1 factoring of M107021149 using up to 24576MB of memory.
[Work thread Oct 29 07:00] Assuming no factors below 2^76 and 4 primality tests saved if a factor is found.
[Work thread Oct 29 07:00] Optimal bounds are B1=1729000, B2=94551000
[Work thread Oct 29 07:00] Chance of finding a factor is an estimated 5.53%
[Work thread Oct 29 07:00]
[Work thread Oct 29 07:00] Using FMA3 FFT length 5760K, Pass1=1536, Pass2=3840, clm=1, 8 threads
[Work thread Oct 29 07:00] D: 330, relative primes: 354, stage 2 primes: 5335022, pair%=92.41
[Work thread Oct 29 07:00] Using 16371MB of memory.
[Work thread Oct 29 07:00] Stage 2 init complete. 848 transforms. Time: 9.439 sec.
[Work thread Oct 29 07:00] M107021149 stage 2 is 0.00% complete.
[Work thread Oct 29 07:32] M107021149 stage 2 is 8.20% complete. Time: 1901.034 sec.
[Work thread Oct 29 08:03] M107021149 stage 2 is 16.51% complete. Time: 1903.386 sec.
[Work thread Oct 29 08:35] M107021149 stage 2 is 24.89% complete. Time: 1901.938 sec.
[Work thread Oct 29 09:07] M107021149 stage 2 is 33.34% complete. Time: 1902.549 sec.[/CODE]Is this the opposite error to the one mentioned by [B][COLOR=Navy]gLauss[/COLOR][/B]?

Prime95 2021-10-29 19:50

[QUOTE=kruoli;591953]Wavefront P-1, In build 5, when day and night memory settings are different, stage 2 cannot be resumed from the larger memory part

Is this the opposite error to the one mentioned by [B][COLOR=Navy]gLauss[/COLOR][/B]?[/QUOTE]

I'm afraid stage 2 really did restart from scratch. Working on a fix.

Aexoden 2021-10-29 21:28

[QUOTE=Aexoden;591788]I'm getting really inconsistent performance on my Threadripper PRO 3955WX (16 cores) running Windows 10 for Workstations. I've only ever used 30.7 builds on this machine, so I don't have any earlier points of comparison (but will certainly test earlier versions if it would matter).[/QUOTE]

Just to update on this, it seems the major problem was having memory installed in the incorrect slots (I'm only using four of the eight channels, and at the time I built the machine, the motherboard manual gave inconsistent advice and apparently I guessed wrong.)

That said, when using only 8 of the available 16 cores, the affinities chosen by Prime95 seem to be suboptimal (which might be expected, given that I think I read somewhere that it's optimized assuming all cores are used). (The default selection on a 5760K CERT run gave about 4.7ms/iter, whereas a manual Affinity setting of 0,4,8,12,16,20,24,28 gave me 3.9ms/iter). I have no idea why that pattern would be better, that was just the fastest one I found.

techn1ciaN 2021-11-02 23:23

Probably a stupid question:

When one is using the cert range controls provided in undoc.txt (Cert[Min|Max]Exponent), do you need to add both lines for the functionality to work, or are you fine to add only one or the other?

Prime95 2021-11-03 03:32

[QUOTE=techn1ciaN;592318]Probably a stupid question:

When one is using the cert range controls provided in undoc.txt (Cert[Min|Max]Exponent), do you need to add both lines for the functionality to work, or are you fine to add only one or the other?[/QUOTE]

I suspect just one will work

sparticus42 2021-11-05 10:48

[QUOTE=techn1ciaN;592318]Probably a stupid question:

When one is using the cert range controls provided in undoc.txt (Cert[Min|Max]Exponent), do you need to add both lines for the functionality to work, or are you fine to add only one or the other?[/QUOTE]


I only have CertMinExponent=0 and have never noticed any issues.


All times are UTC. The time now is 07:44.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2022, Jelsoft Enterprises Ltd.