Prime95 version 25.5

[StarQwest]

Originally Posted by ixfd64

Yes, the L2 cache and physical memory are two different things. I tried increasing my "L2 cache" to 32,768 kb like you described, and nothing happened.

OK, my mistake. I am just trying anything I can find to reduce the per iteration time. It is interesting to see how the new version works when handling such large tests. Also, has anyone completed a 10 million digit LL test on the new version yet to make sure there are no bugs here? I switched my current test of M39,584,257 over to v25.5 and it will finish in about ten hours. I am hoping that it finishes successfully and that there are no bugs. Furthermore, it should be noted that in v24.14 and in v25.5 when using one core, the per iteration time here was 0.101 seconds. When using v25.5 with 2 cores, the per iteration time decreased to 0.053 seconds, a 91% increase in speed, which is much more than the 48% increase from 1 to 2 cores on M332,192,831. Is this a possible bug, or is it only due to the much larger FFT size? If this cannot be corrected, then it will not be practical to ever test numbers in this range.

[Bundu]

Originally Posted by StarQwest

If this cannot be corrected, then it will not be practical to ever test numbers in this range.

By the time the rest of the project reaches those ranges, it's likely that advancements in technology and Geroge's continued optimisation of p95 that those numbers won't seem so bad then

After all a quick look at the benchmarks page tells me an AMD 486 would race though each iteration every 312 seconds on say M40,000,000. I think that works out to..395 years?

[Bundu]

got this today:

[Sep 27 19:53] RESPONSE:
[Sep 27 19:53] pnErrorResult=21
[Sep 27 19:53] pnErrorDetail=Find t_users failed: user_id = 40
[Sep 27 19:53] ==END==
[Sep 27 19:53] 
[Sep 27 19:53] PrimeNet error 21: Invalid user
[Sep 27 19:53] Find t_users failed: user_id = 40
[Sep 27 19:53] Updating computer information on the server
[Sep 27 19:53] URL: http://v5.mersenne.org/v5server/?v=0.95&px=GIMPS&t=uc&g=329e7c3af44367d71a949f60684fd0c4&hg=d136e2271ca5c029d9e5c905cc55590e&wg=d41d8cd98f0
[Sep 27 19:53] RESPONSE:
[Sep 27 19:53] pnErrorResult=21
[Sep 27 19:53] pnErrorDetail=Find t_users failed: user_id = 40
[Sep 27 19:53] ==END==
[Sep 27 19:53] 
[Sep 27 19:53] PrimeNet error 21: Invalid user
[Sep 27 19:53] Find t_users failed: user_id = 40
[Sep 27 19:53] Updating computer information on the server
[Sep 27 19:53] URL: http://v5.mersenne.org/v5server/?v=0.95&px=GIMPS&t=uc&g=329e7c3af44367d71a949f60684fd0c4&hg=d136e2271ca5c029d9e5c905cc55590e&wg=d41d8cd98f0
[Sep 27 19:53] RESPONSE:
[Sep 27 19:53] pnErrorResult=0
[Sep 27 19:53] pnErrorDetail=SUCCESS
[Sep 27 19:53] g=329e7c3af44367d71a949f60684fd0c4
[Sep 27 19:53] u=Team_Bundu
[Sep 27 19:53] un=Team_Bundu
[Sep 27 19:53] cn=bundu4
[Sep 27 19:53] od=15
[Sep 27 19:53] ==END==
[Sep 27 19:53] 
[Sep 27 19:53] Sending expected completion date for F23: Sep 27 2007
[Sep 27 19:53] URL: http://v5.mersenne.org/v5server/?v=0.95&px=GIMPS&t=ap&g=329e7c3af44367d71a949f60684fd0c4&k=AA5AEC51C05284C663CA342DB30893C4&stage=C1S2&c=0&
[Sep 27 19:53] RESPONSE:
[Sep 27 19:53] pnErrorResult=0
[Sep 27 19:53] pnErrorDetail=SUCCESS
[Sep 27 19:53] ==END==
[Sep 27 19:53] 
[Sep 27 19:53] Sending expected completion date for M1074511: Sep 27 2007

[Prime95]

Originally Posted by StarQwest

I am just trying anything I can find to reduce the per iteration time. It is interesting to see how the new version works when handling such large tests. Also, has anyone completed a 10 million digit LL test on the new version yet to make sure there are no bugs here?

Whether you are using version 24.14 or 25.5 you should not attempt this exponent unless you have 2 computers testing the number and compare residues every 100,000 or 1,000,000 iterations. It would be a great waste of time and energy to run just one LL test that is vulnerable to random memory errors, overheating errors, or programmer error.

There have been some double-checks done with 25.5 that matched ealier prime95 results. This does not guarantee the huge FFT you are running has no programming errors.

[Prime95]

Originally Posted by Bundu

got this today:

[Sep 27 19:53] RESPONSE:
[Sep 27 19:53] pnErrorResult=21
[Sep 27 19:53] pnErrorDetail=Find t_users failed: user_id = 40
[Sep 27 19:53] ==END==

Aack, can you send me a few dozen lines prior to this?

[StarQwest]

Originally Posted by Prime95

Whether you are using version 24.14 or 25.5 you should not attempt this exponent unless you have 2 computers testing the number and compare residues every 100,000 or 1,000,000 iterations. It would be a great waste of time and energy to run just one LL test that is vulnerable to random memory errors, overheating errors, or programmer error.

There have been some double-checks done with 25.5 that matched ealier prime95 results. This does not guarantee the huge FFT you are running has no programming errors.

I have currently completed 50,000 iterations and have had no trouble. Also, M100,000,007 finished flawlessly (other than it being composite, which sucked). However, you have a good point about doublechecking it. Yxine doublechecked M100,000,007 while I tested it and we compared interim residues every 1,000,000 iterations. Unfortunately, I do not have another computer available to me, but if anyone else is up to the challenge and wants to doublecheck it, by all means go for it, as it will be worth ~900 CPU years. Also, if anyone has a super fast computer with 4GB of RAM or more and wants to try P-1 stage 2 (I could not- my computer is too slow) or run a few ECM curves that would be really helpful. I have already trial factored to 2^77 and done P-1 stage 1 with B1=B2=4,880,000 and found no factor.

[StarQwest]

Originally Posted by StarQwest

Also, has anyone completed a 10 million digit LL test on the new version yet to make sure there are no bugs here? I switched my current test of M39,584,257 over to v25.5 and it will finish in about ten hours. I am hoping that it finishes successfully and that there are no bugs.

The LL test of M39,584,257 finished fine on v25.5. Apparently there are no bugs for LL testing. The last 3 million iterations of the test performed very well. M332,192,831 is currently performing very well at 0.650 seconds per iteration. Unfortunately however, M39,584,257 is not prime.

[Mr. P-1]

Originally Posted by StarQwest

Also, if anyone has a super fast computer with 4GB of RAM or more and wants to try P-1 stage 2 (I could not- my computer is too slow) or run a few ECM curves that would be really helpful. I have already trial factored to 2^77 and done P-1 stage 1 with B1=B2=4,880,000 and found no factor.

Do you have the save file for the Stage 1 P-1? Prime95 normally deletes these when factoring is complete, so you would have had to copy it just before then.

I just did some tests. It took about 8 times as long to run an ECM curve to comparable limits as it did to do a P-1 test on a 12M exponent. Even if you had software which could do such a test, it probably wouldn't be feasible on your humongous exponent.

[Andi47]

I have just downloaded version 25.5 to a new Core 2 Duo 4300 @1,8 GHz and started a torture test which will run over the weekend. When the PC passes the test, I will join beta testing at the beginning of next week.

[James Heinrich]

Originally Posted by Prime95

Is it dying because of overcommitting memory? I have a fix for that particular bug. I've uploaded a new Win32 version for you to try.

I don't think it's memory-related. It's limited to 128MB and I haven't noticed it exceed that. Your latest win32 version has happily run a little ECM and a lot of TF overnight on that machine, so perhaps you have fixed the bug. I'm starting it on all ECM today to see how it likes it.

[James Heinrich]

Originally Posted by James Heinrich

I don't think it's memory-related. It's limited to 128MB and I haven't noticed it exceed that. Your latest win32 version has happily run a little ECM and a lot of TF overnight on that machine, so perhaps you have fixed the bug. I'm starting it on all ECM today to see how it likes it.

Maybe it's not so happy. No problems with 10 hours of TF, no problems in 14 hours of torture test, but I ran ECM (2 threads) for not much more than an hour and got 5 SUMOUT errors and then finally it crashed outright.

edit:: after a reboot, v25.5 ran for < 5 minutes and crashed out again with

The instruction at "0x0044dded" referenced memory at "0x01480000". The memory could not be "read".