Prime95 30.8 (big P-1 changes, see post #551) - Page 65

LaurV · 2022-09-10, 10:09

Quote:

Originally Posted by axn

Get rid of the TF parameter.

Or add assignment key, or N/A if you have none. P95 has a strange way to count the parameters in assignment lines.

axn · 2022-09-10, 14:24

Quote:

Originally Posted by LaurV

Or add assignment key, or N/A if you have none. P95 has a strange way to count the parameters in assignment lines.

I dont think that would help. This is not an issue related to parameter parsing. Rather, it is intended behavior.

storm5510 · 2022-09-10, 16:42

Maybe a bit off-topic:

Build 16 has no issues running wavefront P-1's that I have seen. Very stable.

preda · 2022-09-11, 09:18

Hello, I have a CPU system with 128GB RAM, and I'd like to have P-1 second-stage use most of that RAM (let's say, 110GB). I see that the work of one worker is split roughly half/half between the first and second stage. This fact (approximate equal duration of first and second stage) opens the possibility to run *two* workers (doing two independent P-1) in parallel, each worker using "all" the RAM for the second stage, by running them such that when one is duing second-stage (requiring the RAM), the other is doing first-stage (low RAM required).

For this to work reliably, mprime (the software) needs to account for this pattern, by "locking" the RAM upon start of the second-stage, and making the other worker that may attempt second-stage wait until the RAM becomes available.

Maybe mprime already does that, and I can use this pattern already?

Cheers!

Prime95 · 2022-09-11, 09:39

Quote:

Originally Posted by preda

Maybe mprime already does that, and I can use this pattern already?

Yes, the setting described in undoc.txt is MaxHighMemWorkers=n the default value for n is one. So, by default mprime setup to run two workers will do what you want.

axn · 2022-09-11, 11:17

Quote:

Originally Posted by preda

Hello, I have a CPU system with 128GB RAM, and I'd like to have P-1 second-stage use most of that RAM (let's say, 110GB). I see that the work of one worker is split roughly half/half between the first and second stage. This fact (approximate equal duration of first and second stage) opens the possibility to run *two* workers (doing two independent P-1) in parallel, each worker using "all" the RAM for the second stage, by running them such that when one is duing second-stage (requiring the RAM), the other is doing first-stage (low RAM required).

I am not sure this offers significantly more throughput (if at all) compared to simple-minded 1-worker-all-core setup for both 1st and 2nd stages. I'd advise you to run both configurations for a few days to see which one is better.

@George: How should Stage2ExtraThreads be set when running multiple workers? When running a single worker, you could just set it to all spare hyperthreads, but not sure how to do it for the multiple worker case.

Prime95 · 2022-09-11, 15:01

Quote:

Originally Posted by axn

@George: How should Stage2ExtraThreads be set when running multiple workers? When running a single worker, you could just set it to all spare hyperthreads, but not sure how to do it for the multiple worker case.

Probably best to set Stage2ExtraThreads to zero. However, experimentation is required to get the true answer.

preda · 2022-09-12, 18:57

Quote:

Originally Posted by Prime95

Yes, the setting described in undoc.txt is MaxHighMemWorkers=n the default value for n is one. So, by default mprime setup to run two workers will do what you want.

Quote:

Originally Posted by axn

I am not sure this offers significantly more throughput (if at all) compared to simple-minded 1-worker-all-core setup for both 1st and 2nd stages. I'd advise you to run both configurations for a few days to see which one is better.

Thank you for the answers!

Now I'm thinking of running only the P-1 second-stage on the high-RAM CPU, and running the corresponding first-stage on a low-RAM GPU (e.g. using gpuowl). This way the valuable resource (lots of RAM) would see maximal usage. For this probably I'd need to generate a P-1 residue (savefile) in the format expected by mprime.

And a little question: would P-1 on the CPU detect bad RAM to any degree? or should I run a PRP/DC from time to time to validate the RAM?

Prime95 · 2022-09-12, 19:25

Quote:

Originally Posted by preda

And a little question: would P-1 on the CPU detect bad RAM to any degree?

P-1 will not catch any RAM errors. A PRP/Gerbicz run or Torture Test would be a good idea.

testcb00 · 2022-09-14, 15:55

I may post my question about Prime95 to the wrong section previously...
https://www.mersenneforum.org/showthread.php?t=28043
https://www.mersenneforum.org/showthread.php?p=613402

I buy an old 1U Supermicro Server recently and I try to use Prime95 to test the stability. I find that there are rounding error exists in FFT 224K/240K when the Memory to use (in MB) > 81919. Lets say 81920MB, the error will exist (both enable/disable AVX). It passed all tests except FFT 224K/240K. It also passed FFT 224K/240K if the Memory to use (in MB) <= 81919.

I am interesting on this behavior, I try the combination of BIOS option but it does not change the range of passed Memory to use (in MB). It keeps 81919MB...

Also, I have tested using another Supermicro E5 v4 server / 128GB DDR4 ECC RDIMM with AVX2 disabled and it passed all test including the FFT 224K/240K.

Is it a defect of CPU/Motherboard? or my motherboard BIOS do not have the updated microcode patch? or it is a bug of Prime95?

Server Details
Supermicro X9SRG-F
Intel Xeon E5-2648Lv2 10C20T
128GB DDR3-1833 RDIMM (Samsung M393B2G70DB0-CMA 16GB x8)

storm5510 · 2022-09-22, 14:59

I found Build 16 stopped during an assignment earlier today. I leave the monitor off most of the time so I didn't see it right away.

It was running an ECM on what appeared to be a wave-front number, 116,466,793. I would think an exponent of this magnitude would be way out-of-bounds for an ECM, but apparently not. The day/night RAM allocation for Stage 2 is 4 GB. I looked at my results on Primenet and it's been running assignments this high for 4 days.

I intend to let it finish then cut the RAM way down. It is possible my favorite pest (cat) stepped on the Esc key. He likes to walk around in places he should not. There is no way to be sure.

2022-09-11, 09:18	#708
preda "Mihai Preda" Apr 2015 2644₈ Posts	Second-stage memory use with dual worker Hello, I have a CPU system with 128GB RAM, and I'd like to have P-1 second-stage use most of that RAM (let's say, 110GB). I see that the work of one worker is split roughly half/half between the first and second stage. This fact (approximate equal duration of first and second stage) opens the possibility to run two workers (doing two independent P-1) in parallel, each worker using "all" the RAM for the second stage, by running them such that when one is duing second-stage (requiring the RAM), the other is doing first-stage (low RAM required). For this to work reliably, mprime (the software) needs to account for this pattern, by "locking" the RAM upon start of the second-stage, and making the other worker that may attempt second-stage wait until the RAM becomes available. Maybe mprime already does that, and I can use this pattern already? Cheers!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Do not post your results here!	kar_bon	Prime Wiki	40	2022-04-03 19:05
what should I post ?	science_man_88	science_man_88	24	2018-10-19 23:00
Where to post job ad?	xilman	Linux	2	2010-12-15 16:39
Moderated Post	kar_bon	Forum Feedback	3	2010-09-28 08:01
Something that I just had to post/buy	dave_0273	Lounge	1	2005-02-27 18:36

2022-09-10, 16:42	#707
storm5510 Random Account Aug 2009 Not U. + S.A. 4737₈ Posts	Maybe a bit off-topic: Build 16 has no issues running wavefront P-1's that I have seen. Very stable.

2022-09-14, 15:55	#714
testcb00 Sep 2022 2×3 Posts	I may post my question about Prime95 to the wrong section previously... https://www.mersenneforum.org/showthread.php?t=28043 https://www.mersenneforum.org/showthread.php?p=613402 I buy an old 1U Supermicro Server recently and I try to use Prime95 to test the stability. I find that there are rounding error exists in FFT 224K/240K when the Memory to use (in MB) > 81919. Lets say 81920MB, the error will exist (both enable/disable AVX). It passed all tests except FFT 224K/240K. It also passed FFT 224K/240K if the Memory to use (in MB) <= 81919. I am interesting on this behavior, I try the combination of BIOS option but it does not change the range of passed Memory to use (in MB). It keeps 81919MB... Also, I have tested using another Supermicro E5 v4 server / 128GB DDR4 ECC RDIMM with AVX2 disabled and it passed all test including the FFT 224K/240K. Is it a defect of CPU/Motherboard? or my motherboard BIOS do not have the updated microcode patch? or it is a bug of Prime95? Server Details Supermicro X9SRG-F Intel Xeon E5-2648Lv2 10C20T 128GB DDR3-1833 RDIMM (Samsung M393B2G70DB0-CMA 16GB x8)

2022-09-22, 14:59	#715
storm5510 Random Account Aug 2009 Not U. + S.A. 7×19² Posts	I found Build 16 stopped during an assignment earlier today. I leave the monitor off most of the time so I didn't see it right away. It was running an ECM on what appeared to be a wave-front number, 116,466,793. I would think an exponent of this magnitude would be way out-of-bounds for an ECM, but apparently not. The day/night RAM allocation for Stage 2 is 4 GB. I looked at my results on Primenet and it's been running assignments this high for 4 days. I intend to let it finish then cut the RAM way down. It is possible my favorite pest (cat) stepped on the Esc key. He likes to walk around in places he should not. There is no way to be sure.