Prime95 v30.4/30.5/30.6

[Prime95]

Originally Posted by LaurV

OTOH, he uses 672k, which should be right for this range (if you look to my last reported results on P+1, I am tickling the 13M expos, took the first 30 or so, for a spin)
Attachment 24774

Fixed in build 4.

Until I upload that version, shy away from AVX-512 FFT lengths that are a multiple of 7.

[Prime95]

Originally Posted by R. Gerbicz

This is still doesn't find q in the first test.
More greedy run, try:
Pplus1=N/A,1,2,86743,-1,86742,2000000,1,40
Pplus1=N/A,1,2,86743,-1,86743,2000000,1,40.

Including the Mersenne exponent in the P+1 bounds is cheap, but also encourages using the wrong tool to find P-1 factors.

1) P-1 factoring is faster than P+1 factoring.
2) P-1 does not miss any B1/B2 smooth factor-1 values, whereas P+1 factoring misses half of them.

Unless someone points to a good reason to use P+1 factoring to find P-1 factors, I'm not inclined to add the Mersenne exponent to the B1 bound.

[Prime95]

Originally Posted by tha

Downloaded the new version. It now does not report anything about the composite factor reported by the client to the server. I would expect a list of factors already known to the server and, if present, a new factor as a new factor.

I must admit I have not found a new factor with the new version so I continue to do P-1 factoring on exponents with three factors or more already found and no or little previous P-1 done before. Will report when I have found one.

The fix I made was to no longer truncate the message you were getting:

[Comm thread Apr 22 18:59] Already have factor 72324517147464481 for M9994027
[Comm thread Apr 22 18:59] Already

When you say "now does not report anything", what message did you get and what were you expecting?

[Prime95]

Originally Posted by kriesel

But an even larger exponent launches ok in Pminus1 worktodo form on the 7250 on a 42M fft

I tried your problem exponent on my laptop with version 30.6b3. It launched OK using a 36M FFT.

I don't know what is going on.

[Prime95]

Originally Posted by gLauss

I am wondering what the optimal P-1/P+1/ECM strategy is and what P+1 bounds I should choose for my exponents. My current strategy is:
1) Get current ECM level, e.g. B1=3M.
2) George said that the P+1 should be done to two stages above the current ECM level, e.g. B1=50M.
3) Before the P+1 is done, one should do a P-1 run with at least twice the bound of the P+1, e.g. B1=100M.
4) If there is no ECM done, then I think it is optimal to do choose B1 for the P+1 as the current P-1 B1 divided by a factor of 2 to 5.
However, this is only my gutfeeling. Is there someone who can calculate the odds?

As to "George said", more accurate would be "George suggested". Like you, I'm going on a gut feeling.

I do not think there is an optimal strategy. I can say without a doubt that the right order is 1) try P-1, 2) try P+1, 3) try ECM. You want to choose what are seemingly excessive P-1 and P+1 bounds because if you ever decide to increase the bounds you have to re-do all the work that was originally done.

[axn]

Originally Posted by Prime95

Including the Mersenne exponent in the P+1 bounds is cheap, but also encourages using the wrong tool to find P-1 factors.

1) P-1 factoring is faster than P+1 factoring.
2) P-1 does not miss any B1/B2 smooth factor-1 values, whereas P+1 factoring misses half of them.

Unless someone points to a good reason to use P+1 factoring to find P-1 factors, I'm not inclined to add the Mersenne exponent to the B1 bound.

Noone is asking about using P+1 runs instead of P-1 runs to find P-1 factors. But, sometimes machines makes mistakes and P-1 runs could miss factors. If we're running P+1 anyways, and if it is capable of finding some P-1 factors, and sometimes P-1 factors are missed by prior runs, it would be good to let P+1 have the capability to find some of them.

[Falkentyne]

Originally Posted by Prime95

As to "George said", more accurate would be "George suggested". Like you, I'm going on a gut feeling.

I do not think there is an optimal strategy. I can say without a doubt that the right order is 1) try P-1, 2) try P+1, 3) try ECM. You want to choose what are seemingly excessive P-1 and P+1 bounds because if you ever decide to increase the bounds you have to re-do all the work that was originally done.

Hi, I didn't test 30.6 beta 4 but beta 2 and 3 are broken on Rocket Lake (11900k).
testing large FFT with AVX disabled causes all threads to crash at the same time at the end of the first iteration.

30.5 beta 2 works fine.

[tha]

Originally Posted by Prime95

The fix I made was to no longer truncate the message you were getting:

[Comm thread Apr 22 18:59] Already have factor 72324517147464481 for M9994027
[Comm thread Apr 22 18:59] Already

When you say "now does not report anything", what message did you get and what were you expecting?

I did find a new factor this morning, and indeed it does no longer truncate the message:

PrimeNet success code with additional info:
[Comm thread Apr 28 05:52] Composite factor 1433220254355182327305060291296816494851568454069468578543 = 78687058243775872759827977 * 3027792433 * 5563081984303 * 1081354441
[Comm thread Apr 28 05:52] Already have factor 3027792433 for M9011287
[Comm thread Apr 28 05:52] Already have factor 5563081984303 for M9011287
[Comm thread Apr 28 05:52] Already have factor 1081354441 for M9011287

What I was expecting was that each composite factor reported to the server would still be factorized in the return message. But indeed that does not serve any real purpose. The message 'factoring result was not needed' is fine.

One more request though, for mprime, if one chooses option 4, 'Test/Continue' the first output to the screen is a re-display of the menu. That does not serve any purpose as it is the last on display and the second one only clutters the screen. Since nowadays all terminal screens are scroll-able it would make more sense to have the re-display of the menu following a ^C.

[R. Gerbicz]

Originally Posted by axn

Noone is asking about using P+1 runs instead of P-1 runs to find P-1 factors. But, sometimes machines makes mistakes and P-1 runs could miss factors. If we're running P+1 anyways, and if it is capable of finding some P-1 factors, and sometimes P-1 factors are missed by prior runs, it would be good to let P+1 have the capability to find some of them.

Exactly, add to that there is no error check in P-1. You could get a partial error check in stage one if you are doing a mixed prp/pm1 test, what gpuowl is doing. I could be wrong but you could also get error in gpuowl's P-1 first stage run [this is an error in P-1, not in the prp part].

And what would be the optimal run for a mixed prp/pm1 test?
If a full P+1 test before this [obviously not with gpuowl, this is not supported there] then you lost many factors, need to halt the prp test after the p-1 run. It is possible that the optimal run is using increasing B1,B2 bounds for P+-1 runs.

[kriesel]

Originally Posted by Prime95

I tried your problem exponent on my laptop with version 30.6b3. It launched OK using a 36M FFT.

I don't know what is going on.

The plot thickens. The Xeon Phi 7210 could run the M660M P-1, until it couldn't, for unknown reason. Excerpted from the worker window:

[Apr 27 11:12] Worker starting
[Apr 27 11:12] Setting affinity to run worker on CPU core #1
[Apr 27 11:12] 
[Apr 27 11:12] P-1 on M660000031 with B1=3600000, B2=180000000
[Apr 27 11:12] Setting affinity to run helper thread 1 on CPU core #2
[Apr 27 11:12] Setting affinity to run helper thread 2 on CPU core #3
[Apr 27 11:12] Setting affinity to run helper thread 7 on CPU core #8
[Apr 27 11:12] Setting affinity to run helper thread 4 on CPU core #5
[Apr 27 11:12] Setting affinity to run helper thread 5 on CPU core #6
[Apr 27 11:12] Using AVX-512 FFT length 36M, Pass1=1536, Pass2=24K, clm=2, 16 threads
[Apr 27 11:12] Setting affinity to run helper thread 3 on CPU core #4
[Apr 27 11:12] Setting affinity to run helper thread 8 on CPU core #9
[Apr 27 11:12] Setting affinity to run helper thread 9 on CPU core #10
[Apr 27 11:12] Setting affinity to run helper thread 10 on CPU core #11
[Apr 27 11:12] Setting affinity to run helper thread 11 on CPU core #12
[Apr 27 11:12] Setting affinity to run helper thread 12 on CPU core #13
[Apr 27 11:12] Setting affinity to run helper thread 13 on CPU core #14
[Apr 27 11:12] Setting affinity to run helper thread 14 on CPU core #15
[Apr 27 11:12] Setting affinity to run helper thread 6 on CPU core #7
[Apr 27 11:12] Setting affinity to run helper thread 15 on CPU core #16
[Apr 27 11:30] M660000031 stage 1 is 0.19% complete. Time: 1109.194 sec.
...
[Apr 27 22:59] M660000031 stage 1 is 7.70% complete. Time: 1049.200 sec.
[Apr 27 23:17] M660000031 stage 1 is 7.89% complete. Time: 1058.078 sec.
[Apr 27 23:19] Worker stopped while running needed benchmarks.
[Apr 27 23:26] Benchmarks complete, restarting worker.
[Apr 27 23:26] 
[Apr 27 23:26] P-1 on M660000031 with B1=3600000, B2=180000000
[Apr 27 23:26] Cannot initialize FFT code, errcode=1002
[Apr 27 23:26] Worker stopped.

There's nothing about the issue in prime.log. It gets logged in results.txt but there is no additional information

[Tue Apr 27 23:26:58 2021]
Cannot initialize FFT code, errcode=1002

Is there some debugging I could turn on for this? Searching for "debug" in the documentation files revealed only the unrelated following in whatsnew.txt (which maybe ought be edited to remove Scott Kurowski's name if he's no longer actively participating):

New features in Version 16.5 of prime95.exe
-------------------------------------------

1)  A new httpnet.dll can provide Scott Kurowski with debugging information.
    Simply create a primenet.ini file with these lines:
    [Debug]
    PacketLog=1
    Output file is pnHttp.txt.

Similarly searching for "verbos" yielded only in undoc.txt,

You can change the amount of information the program outputs while setting affinities.
    AffinityVerbosity=0 or 1    (default is 1)
    AffinityVerbosityTorture=0 or 1    (default is 0)
    AffinityVerbosityTime=0 or 1    (default is 0)
    AffinityVerbosityBench=0 or 1    (default is 0)
These settings are made in prime.txt.  Zero for less output, one for more output.
The first setting is for all normal work (LL, TF, ECM, P-1, P+1, PRP).  The others
are for torture testing, Advanced/Time, and benchmarking.

Also, although there was other work in the worker's section of the worktodo, it stopped dead and left 16 cores idle for hours, until found.
It would be good if there was an option to have it comment out a problem worktodo line, and attempt the next in line, to keep it productive. In this case that worktodo section was (with AID redaction):

[Worker #1]
Pminus1=N/A,1,2,660000031,-1,3600000,180000000
PRP=(aid),1,2,350021701,-1,79,0

so it could have progressed to a different work type, different exponent, different fft length and probably remained productive, by modifying the relevant worktodo section to something like the following and then restarting the worker:

[Worker #1]
;Pminus1=N/A,1,2,660000031,-1,3600000,180000000
;preceding commented out because [Apr 27 23:26] Cannot initialize FFT code, errcode=1002
PRP=(aid),1,2,350021701,-1,79,0

To recap:
It interrupts work it is doing, to do more benchmarks;
it indicates there's an issue to resume the interrupted work in progress;
it doesn't tell WHY it can no longer do what it had previously been doing before interrupted by benchmarking;
there does not appear to be any option for having it tell more about the problem it encountered preventing resumption;
it has other work it could switch to, to avoid loss of throughput, but does not support switching to that automatically.
I had hoped to do P-1 of large exponents on FMA3 and AVX512, but have now found that 2 of the 4 AVX512 (both Xeon Phi, different models) have some undetermined issues with it, unpredictably. Attempting them on the FMA3 i7-8750H or AVX512 i5-1035G1 laptops would take weeks. I may try continuing the ~8% complete M660000031 there anyway, rather than throwing away the progress and running it on gpuowl / Radeon VII (which are deeply queued already).

If a prerelease prime95 version with greater P-1 or fft startup debugging output became available (in forum thread, or by dropbox link in PM?) I would be happy to try it out.


update: An AVX512 i5-1035g1 laptop prime95 v30.6b2, Windows 10 build 20H2 19042.928, 16gb ram) was able to resume the M660M P-1 run in progress, with ETA 3.5 weeks. We'll see how that goes.

[Prime95]

Originally Posted by kriesel

The plot thickens.

Try renaming gwnum.txt. Restart prime95, will the exponent now run? If so, email/PM the gwnum.txt file to me. Thanks.