Prime95 30.8 (beta, big P-1 changes, see post #551) - Page 26

Prime95 · 2022-01-01, 23:54

30.8 build 7

A few bugs were hopefully fixed.

This build concentrates on improving multithreading when working on exponents below say 1M. The gwnum library does not support multithreading for one-pass FFTs and does not scale particularly well for the smallest two-pass FFTs. In the previous build polymult was fully multithreaded, but the poly preparation and poly results processing relied on gwnum library for any multithreading. For my work on exponents around M80000, stage 2 was spending about half the time quad-threaded and half the time single-threaded. Much better now -- about 35% faster.

The default is to not use gwnum's multithreading for FFT sizes below 256K.

The problem with severely over-allocating memory for small exponents (below say 40K) is a little better, but by no means fixed. I understand the issue, but do not yet understand the best solution.

For those that want to P-1 exponents below 100K (and maybe below 1M or 2M), I do the following on my quad core box. Build a big worktodo file using 4 workers, 1 core per worker. Each worktodo line does stage 1 only. Example:

Code:

Pminus1=1,2,79039,-1,1000000000,1000000000

When that work completes, I switch to 1 worker 4 cores per worker and all the RAM I can afford. Worktodo lines run stage 2, example:

Code:

Pminus1=1,2,79039,-1,1000000000,1000000000,117

Windows 64-bit: https://mersenne.org/ftp_root/gimps/p95v308b7.win64.zip
Linux 64-bit: https://mersenne.org/ftp_root/gimps/...linux64.tar.gz

SethTro · 2022-01-02, 00:25

Quote:

Originally Posted by Prime95

This would be a different issue. Can you provide more details? What appears on the screen? Can you provide save file, worktodo.txt, prime.txt, local.txt?

I'm no longer able to reproduce so I don't know what to say. With three workers, with one worker, on build 6 and build 7 I can get this to entering stage 2. I have included the logs (and local.txt and prime.txt) from build 6 when it was failing to finish.

The setup when this was failing was 3 workers, with 8 cores for the last worker.
mprime would consistently run with 100% cpu on the 8 cores then all of those cores would drop to 0% without any status lines (or without saving a finished stage 1 file) and mprime would be stuck such that Ctrl+C would just hangs (repeatedly printing "Waiting for worker threads to stop.")

The work item was

Code:

Pminus1=1,2,38197477,-1,25000000,0,"229184863,1527899081,16730494927,1442642311337,2675886053759,2439808457033351,31168004071948307743,470793337824880328249,2381826327823260512809"

Here's some old screen logs (from build 6)

Code:

[Worker #3 Dec 30 00:40] M38197477 stage 1 is 99.46% complete.
[Worker #2 Dec 30 00:40] M9100919 stage 1 is 47.34% complete.
[Worker #1 Dec 30 01:27] M1938317 stage 1 is 38.85% complete. Time: 2792.084 sec.
[Worker #1 Dec 30 02:13] M1938317 stage 1 is 39.29% complete. Time: 2788.538 sec.
[Worker #2 Dec 30 02:19] M9100919 stage 1 is 48.72% complete. Time: 5911.293 sec.


             Invalid choice

             Main Menu

         1.  Test/Primenet
         2.  Test/Workers
         3.  Test/Status
         4.  Test/Stop
         5.  Test/Exit
         6.  Advanced/Test
         7.  Advanced/Time
         8.  Advanced/P-1
         9.  Advanced/ECM
        10.  Advanced/Manual Communication
        11.  Advanced/Unreserve Exponent
        12.  Advanced/Quit Gimps
        13.  Options/CPU
        14.  Options/Resource Limits
        15.  Options/Preferences
        16.  Options/Torture Test
        17.  Options/Benchmark
        18.  Help/About
        19.  Help/About PrimeNet Server
Your choice: 5

[Main thread Dec 30 02:23] Stopping all worker windows.
Waiting for worker threads to stop.
[Worker #2 Dec 30 02:23] Worker stopped.
^C[Main thread Dec 30 02:23] Stopping all worker windows.
^Z
[1]+  Stopped                 ./mprime -m -d
seven@seven:~/Projects/GIMPS/p95v308b6$ kill %1

CPU usage had dropped to 0 without printing anything more about thread 3.

I repeated this 3 times and the output is all the same (attempt 3 here) giving it plenty of time to complete. This final .5% should take roughly 300 seconds which is < 10 minutes or the 2 hours in the first log.

Code:

[Worker #3 Dec 30 02:36] Worker starting
[Worker #3 Dec 30 02:36] Setting affinity to run worker on CPU core #6
[Worker #3 Dec 30 02:36]
[Worker #3 Dec 30 02:36] P-1 on M38197477 with B1=25000000, B2=2500000000
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 1 on CPU core #7
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 2 on CPU core #8
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 5 on CPU core #11
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 4 on CPU core #10
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 6 on CPU core #12
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 7 on CPU core #13
[Worker #3 Dec 30 02:36] Setting affinity to run helper thread 3 on CPU core #9
[Worker #3 Dec 30 02:36] Using AVX FFT length 2M, Pass1=512, Pass2=4K, clm=4, 8 threads
[Worker #3 Dec 30 02:36] M38197477 stage 1 is 99.46% complete.
[Worker #2 Dec 30 02:36] M9100919 stage 1 is 48.94% complete.


             Invalid choice

             Main Menu

         1.  Test/Primenet
         2.  Test/Workers
         3.  Test/Status
         4.  Test/Stop
         5.  Test/Exit
         6.  Advanced/Test
         7.  Advanced/Time
         8.  Advanced/P-1
         9.  Advanced/ECM
        10.  Advanced/Manual Communication
        11.  Advanced/Unreserve Exponent
        12.  Advanced/Quit Gimps
        13.  Options/CPU
        14.  Options/Resource Limits
        15.  Options/Preferences
        16.  Options/Torture Test
        17.  Options/Benchmark
        18.  Help/About
        19.  Help/About PrimeNet Server
Your choice: 5

[Main thread Dec 30 02:46] Stopping all worker windows.
Waiting for worker threads to stop.
[Worker #2 Dec 30 02:46] Worker stopped.
[Worker #1 Dec 30 02:46] Worker stopped.

^C[Main thread Dec 30 02:46] Stopping all worker windows.
^Z
[1]+  Stopped                 ./mprime -m -d

axn · 2022-01-02, 03:30

Build 7 observations:

Extra helper thread affinity works properly.
Init is 10% faster (not a super big poly - presumably bigger ones will benefit more).
Using smaller FFT size for stage 2 (round offs are, consequently, bigger)
More details included in results.json, presumably allowing proper credit calculation.

Consequently, Manual submission form chokes on the new JSON format

EDIT:- B2 adjustment now works for fixed B2 as well

James Heinrich · 2022-01-02, 04:17

Quote:

Originally Posted by axn

Manual submission form chokes on the new JSON format

Should be fixed now, please try again.

Luminescence · 2022-01-02, 18:26

One of my workers just spit out this line while Stopping/exiting during stage 2 in build 7:

Code:

5

[Main thread Jan 2 19:22] Stopping all worker windows.
Waiting for worker threads to stop.
Waiting for worker threads to stop.
Waiting for worker threads to stop.
Waiting for worker threads to stop.
[Main thread Jan 2 19:22] In write_giant, unexpected len == 0 failure
[Work thread Jan 2 19:22] Worker stopped.
[Main thread Jan 2 19:22] Execution halted.
[Main thread Jan 2 19:22] Choose Test/Continue to restart.

I'm running a bare server version of Ubuntu 20.04.
Another edit: just had a round-off error with this AVX-512 machine. Exponent is a 9.8M one, could explain why this machine is sometimes going so slow today. I backscrolled a bit before the error, to the init, and noticed this absolute madness:

Code:

[Worker #2 Jan 2 19:46] Using AVX-512 FFT length 504K, Pass1=1344, Pass2=384, clm=1, 9 threads
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:46] M9801149 stage 1 complete. 0 transforms. Total time: 0.000 sec.
[Worker #2 Jan 2 19:46] Round off: 0
[Worker #2 Jan 2 19:46] Inversion of stage 1 result complete. 5 transforms, 1 modular inverse. Time: 2.196 sec.
[Worker #2 Jan 2 19:46] Available memory is 112634MB.
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 1 on CPU core #11
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 5 on CPU core #15
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 2 on CPU core #12
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:46] Switching to AVX-512 FFT length 512K, Pass1=1K, Pass2=512, clm=1, 9 threads
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 4 on CPU core #14
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 3 on CPU core #13
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 7 on CPU core #17
[Worker #2 Jan 2 19:46] Available memory is 112634MB.
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 1 on CPU core #11
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 2 on CPU core #12
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 5 on CPU core #15
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:46] Switching to AVX-512 FFT length 560K, Pass1=896, Pass2=640, clm=1, 9 threads
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 4 on CPU core #14
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 3 on CPU core #13
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 7 on CPU core #17
[Worker #2 Jan 2 19:46] Available memory is 112634MB.
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 1 on CPU core #11
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 2 on CPU core #12
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 5 on CPU core #15
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:46] Switching to AVX-512 FFT length 576K, Pass1=1152, Pass2=512, clm=1, 9 threads
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 4 on CPU core #14
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 3 on CPU core #13
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 7 on CPU core #17
[Worker #2 Jan 2 19:46] Available memory is 112634MB.
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 1 on CPU core #11
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 2 on CPU core #12
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 5 on CPU core #15
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 7 on CPU core #17
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 4 on CPU core #14
[Worker #2 Jan 2 19:46] Switching to AVX-512 FFT length 588K, Pass1=1344, Pass2=448, clm=1, 9 threads
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:46] Setting affinity to run helper thread 3 on CPU core #13
[Worker #2 Jan 2 19:46] Available memory is 112634MB.
[Worker #2 Jan 2 19:46] Using 112635MB of memory.  D: 43890, 4320x19972 polynomial multiplication.

and shortly afterwards:

Code:

[Worker #2 Jan 2 19:50] Round off: 0.2715111504
[Worker #1 Jan 2 19:50] M9817169 stage 1 is 0.69% complete. Time: 5.011 sec.
[Worker #1 Jan 2 19:50] M9817169 stage 1 is 1.38% complete. Time: 4.945 sec.
[Worker #1 Jan 2 19:50] M9817169 stage 1 is 2.07% complete. Time: 4.947 sec.
[Worker #2 Jan 2 19:50] Possible roundoff error (0.46648509), backtracking to last save file and using larger FFT.
[Worker #1 Jan 2 19:50] M9817169 stage 1 is 2.77% complete. Time: 4.772 sec.
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 1 on CPU core #11
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 5 on CPU core #15
[Worker #2 Jan 2 19:50] Using AVX-512 FFT length 512K, Pass1=1K, Pass2=512, clm=1, 9 threads
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 6 on CPU core #16
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 2 on CPU core #12
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 4 on CPU core #14
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 3 on CPU core #13
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 8 on CPU core #18
[Worker #2 Jan 2 19:50] Setting affinity to run helper thread 7 on CPU core #17
[Worker #1 Jan 2 19:50] Restarting worker with new memory settings.
[Worker #2 Jan 2 19:50] M9801149 stage 1 complete. 0 transforms. Total time: 0.000 sec.
[Worker #2 Jan 2 19:50] Round off: 0

kruoli · 2022-01-04, 10:48

Situation:
I wanted to continue my work in the 200k range with the now possible higher bounds.

Problems:

When continuing a stage 1 file that has already completed stage 2 once (in my case B1 = 30M, B2 = 3G), the worktodo entry's B2 must exceed the previous B2 even if B2 shall be automatically computed (by factored_bits).
In the above case, a calculated B2 is not displayed, even though it states B2=TBD. Instead, it choses my B2 from the worktodo (in my case B2 = 5G), although factored_bits was given.

Observations:

For a 200k exponent, b7 chose a way smaller D (b6: > 800000, b7: 2310).
It switched from 10k, 8 threads to 12k 6 threads.
The last step (92.97% to 100%) took way (at least 3x) longer than 70.65% to 92.97%. So 0% to 92.97% took around the same time as 92.97% to 100%. In taskmgr, memory was slowly decreasing during this time.
I am not sure why round-off checking was disabled. I thought it was enabled, but either I disabled it myself and forgot about it or it got disabled by itself. I hope this should not effect what I saw.

Shortened screen output:

Code:

[Work thread Jan 4 11:26] P-1 on M201557 with B1=30000000, B2=TBD
[...]
[Work thread Jan 4 11:29] Using FMA3 FFT length 10K, Pass1=128, Pass2=80, clm=2, 8 threads
[...]
[Work thread Jan 4 11:29] Switching to FMA3 FFT length 12K, Pass1=256, Pass2=48, clm=1, 6 threads
[Work thread Jan 4 11:29] Using 45875MB of memory.  D: 2310, 240x483449 polynomial multiplication.
[...]
[Work thread Jan 4 11:29] Round off: 0, poly_size: 2, EB: 9.5245, SM: 0
[Work thread Jan 4 11:29] Round off: 0, poly_size: 4
[Work thread Jan 4 11:29] Round off: 0, poly_size: 8
[Work thread Jan 4 11:29] Round off: 0, poly_size: 16
[Work thread Jan 4 11:29] Round off: 0, poly_size: 32
[Work thread Jan 4 11:29] Round off: 0, poly_size: 64
[Work thread Jan 4 11:29] Round off: 0, poly_size: 128
[Work thread Jan 4 11:29] Round off: 0, poly_size: 256
[Work thread Jan 4 11:32] Stage 2 init complete. 6089 transforms. Time: 151.955 sec.
[Work thread Jan 4 11:32] Round off: 0
[Work thread Jan 4 11:32] M201557 stage 2 at B2=1301604150 [26.03%]
[Work thread Jan 4 11:33] M201557 stage 2 at B2=2417264850 [48.34%].  Time: 57.183 sec.
[Work thread Jan 4 11:34] M201557 stage 2 at B2=3532925550 [70.65%].  Time: 57.348 sec.
[Work thread Jan 4 11:35] M201557 stage 2 at B2=4648586250 [92.97%].  Time: 57.274 sec.
[Work thread Jan 4 11:38] M201557 stage 2 complete. 19328626 transforms. Total time: 360.524 sec.
[Work thread Jan 4 11:38] Stage 2 GCD complete. Time: 0.110 sec.
[Work thread Jan 4 11:38] M201557 completed P-1, B1=30000000, B2=5764246950, Wi4: 5C1636F6

kruoli · 2022-01-04, 11:38

Another observation, contrary to axn's:

Code:

[Work thread Jan 4 11:38] P-1 on M29833387 with B1=1750000, B2=400000000
[...]
[Work thread Jan 4 11:38] Using FMA3 FFT length 1536K, Pass1=1536, Pass2=1K, clm=1, 8 threads
[stage 1, ...]
[Work thread Jan 4 12:08] Switching to FMA3 FFT length 1600K, Pass1=320, Pass2=5K, clm=2, 8 threads
[...]
[Work thread Jan 4 12:08] Switching to FMA3 FFT length 1680K, Pass1=448, Pass2=3840, clm=2, 8 threads
[...]
[Work thread Jan 4 12:08] Switching to FMA3 FFT length 1792K, Pass1=1792, Pass2=1K, clm=1, 8 threads

For this exponent, the FFT size got increased thrice!

axn · 2022-01-04, 12:25

Quote:

Originally Posted by kruoli

For this exponent, the FFT size got increased thrice!

I think this is still consistent with my experience. My report of smaller FFT size was relative to the previous version.

For 27.0M exponent, stage 1 uses 1536K and stage 2 uses 1680K with build 6 and earlier. However, with build 7, it switched to a 1600K FFT for stage 2.

Looks like it _tries_ to do the same with 29.8M also. It tries to use 1600K FFT, but sees that the round off is bad, increases to 1680K, and finally settles on 1792K. My guess is previous builds would directly jump to 1792K (or higher) FFT. So the general observation is build 7 uses same or smaller FFT for stage 2 compared to previous builds.

firejuggler · 2022-01-04, 16:02

on 30.8b6, a 8.5M exponent get a 448k FFT for stage1 and 512k for stage 2.
I have yet to test with b7

Prime95 · 2022-01-04, 17:54

Quote:

Originally Posted by axn

My guess is previous builds would directly jump to 1792K (or higher) FFT. So the general observation is build 7 uses same or smaller FFT for stage 2 compared to previous builds.

That is correct. I ought to be able to figure out the best FFT size without all that switching. So much to do, so little time...

Prime95 · 2022-01-04, 20:33

Quote:

Originally Posted by kruoli

Situation:
I wanted to continue my work in the 200k range with the now possible higher bounds.

Looks like you'll have to wait until next build.

Quote:

When continuing a stage 1 file that has already completed stage 2 once (in my case B1 = 30M, B2 = 3G), the worktodo entry's B2 must exceed the previous B2 even if B2 shall be automatically computed (by factored_bits).

I'm not sure automatic computation of B2 is coded to handle this case or even if it can be coded to handle it. I'll investigate. If it is at all hard, I'll require worktodo.txt to have an explicit B2.

Quote:

In the above case, a calculated B2 is not displayed, even though it states B2=TBD. Instead, it choses my B2 from the worktodo (in my case B2 = 5G), although factored_bits was given.

Displaying TBD instead of 5G is clearly a bug.

Quote:

[*]For a 200k exponent, b7 chose a way smaller D (b6: > 800000, b7: 2310).[*]It switched from 10k, 8 threads to 12k 6 threads.

How bizarre. At this point I have no ideas. The 8 to 6 thread switch is really weird.

Quote:

[*]The last step (92.97% to 100%) took way (at least 3x) longer than 70.65% to 92.97%. So 0% to 92.97% took around the same time as 92.97% to 100%.

My test case is one of my 80K runs B1=1B, B2=20T. I tried a worktodo of 1B,22T and replicated the problem. The code successfully figures out how many iterations are required to go from 20T to 22T, but starts from 15T rather that the already completed 20T. When prime95 finished reached 100% of the iterations completed it discovered it was way short of the 22T target and continued on until it reached the 22T target.

This bug should not be hard to fix.

2022-01-01, 23:54	#276
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1111011010000₂ Posts	30.8 build 7 A few bugs were hopefully fixed. This build concentrates on improving multithreading when working on exponents below say 1M. The gwnum library does not support multithreading for one-pass FFTs and does not scale particularly well for the smallest two-pass FFTs. In the previous build polymult was fully multithreaded, but the poly preparation and poly results processing relied on gwnum library for any multithreading. For my work on exponents around M80000, stage 2 was spending about half the time quad-threaded and half the time single-threaded. Much better now -- about 35% faster. The default is to not use gwnum's multithreading for FFT sizes below 256K. The problem with severely over-allocating memory for small exponents (below say 40K) is a little better, but by no means fixed. I understand the issue, but do not yet understand the best solution. For those that want to P-1 exponents below 100K (and maybe below 1M or 2M), I do the following on my quad core box. Build a big worktodo file using 4 workers, 1 core per worker. Each worktodo line does stage 1 only. Example: Code: Pminus1=1,2,79039,-1,1000000000,1000000000 When that work completes, I switch to 1 worker 4 cores per worker and all the RAM I can afford. Worktodo lines run stage 2, example: Code: Pminus1=1,2,79039,-1,1000000000,1000000000,117 Windows 64-bit: https://mersenne.org/ftp_root/gimps/p95v308b7.win64.zip Linux 64-bit: https://mersenne.org/ftp_root/gimps/...linux64.tar.gz

2022-01-02, 03:30	#278
axn Jun 2003 5365₁₀ Posts	Build 7 observations: Extra helper thread affinity works properly. Init is 10% faster (not a super big poly - presumably bigger ones will benefit more). Using smaller FFT size for stage 2 (round offs are, consequently, bigger) More details included in results.json, presumably allowing proper credit calculation. Consequently, Manual submission form chokes on the new JSON format EDIT:- B2 adjustment now works for fixed B2 as well Last fiddled with by axn on 2022-01-02 at 03:32

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Prime95 beta version 28.4	Prime95	Software	20	2014-03-02 02:51
Prime95 beta version 28.3	Prime95	Software	68	2014-02-23 05:42
Prime95 version 27.1 early preview, not-even-close-to-beta release	Prime95	Software	126	2012-02-09 16:17
Beta version 24.12 available	Prime95	Software	33	2005-06-14 13:19
Beta version of PRP	Prime95	PSearch	15	2004-09-17 19:21

2022-01-04, 16:02	#284
firejuggler "Vincent" Apr 2010 Over the rainbow 17×167 Posts	on 30.8b6, a 8.5M exponent get a 448k FFT for stage1 and 512k for stage 2. I have yet to test with b7