Trial division with CUDA (mmff) -- used, but runs like new! - Page 25

RichD · 2012-10-24, 01:34

Quote:

Originally Posted by RichD

I've had to restart my mmff-0.26 run on MMFactor=127 but it doesn't appear to look for a .ckp file even though I have checkpoint turned on in the .ini file.

Luckily I spotted this as soon as it started and quickly renamed the .ckp file before quitting. Looking back through this thread I see I can use the -nocheck switch. It seems to be continuing OK. I first thought I may have lost 5-6 days...

aketilander · 2012-11-03, 09:01

The TFs covered by mmff are so large that there will never ever be possible to do a LL to check for compositeness. Therefore I would guess that someone would sooner or later want to do double checks at least when the double mersennes are concerned. Would it be possible to add a checksum or a residue to mmff? Maybe a summarized residue of all the TFs done in a region?

I don't think that mmff has a test in the beginning so sometimes, when you don't find factors you start to distrust your system. Did I OC too much? Are there hardware errors? In that case it would be nice to be able to DC a couple of regions just to make sure that your system works OK.

So, would it be possible and meaningful to add some kind of checksum or residue to mmff?

pinhodecarlos · 2012-11-03, 14:05

I have a question, will mmff client detect the number of GPU's available or not? I mean, should the number of mmff clients be started as many times as the number of GPU cards installed?

Thank you in advance,

Carlos

flashjh · 2012-11-03, 15:01

Quote:

Originally Posted by pinhodecarlos

I have a question, will mmff client detect the number of GPU's available or not? I mean, should the number of mmff clients be started as many times as the number of GPU cards installed?

Thank you in advance,

Carlos

You can run mmff on multiple GPUs. Use the -d 0 or -d 1, etc.

Prime95 · 2012-11-03, 16:21

Quote:

Originally Posted by aketilander

Would it be possible to add a checksum or a residue to mmff?

I don't see how to do this. Even if the second run used the same GPUSievePrimes value, the two runs would test a slightly different set of factors. This is because there are race conditions in the GPU sieve (it is faster to use non-atomic bit operations).

akruppa · 2012-11-07, 09:13

Is there any way to get statistics on the error rate of the trial divison? E.g., what approximate fraction of factors are missed due to race conditions, or due to consumer-grade cards skipping a beat, etc.

Prime95 · 2012-11-07, 16:09

Quote:

Originally Posted by akruppa

Is there any way to get statistics on the error rate of the trial divison? E.g., what approximate fraction of factors are missed due to race conditions, or due to consumer-grade cards skipping a beat, etc.

I have no statistics on the error rate. The program does check two of the trial divisions in each class. So, a couple of thousand are double-checked in each bit-level. Fortunately, there have been few reports of validation failures.

However, the trial factors that are double-checked always come from thread id zero (of 256). If the CUDA scheduler always assigns this thread_id to specific CUDA cores, then a large number of CUDA cores are not getting any double-checking.

As to race conditions, there shouldn't be any problems. The program intentionally has race conditions, but these only cause a few extra composite factors to be tested.

aketilander · 2012-11-16, 08:03

I am wondering about the performance of different videocards. If I compare:

http://www.videocardbenchmark.net/high_end_gpus.html

with James'

http://www.mersenne.ca/mfaktc.php

there seems to be large differences. Lets take GTX 680 as an example. According to the first site its on the top of the list, but according to James' its not so good. I can see that James' lists someting called "Compute 3.0". I am not sure if this refers to the speed of the PCI-e bus or if this is something else. What I am really wondering is if there is an unused potential of some videocards which may be better used in the future when mmff is better adapted to "Compute 3.0" or maybe something else?

axn · 2012-11-16, 08:11

Quote:

Originally Posted by aketilander

I am wondering about the performance of different videocards. If I compare:

http://www.videocardbenchmark.net/high_end_gpus.html

with James'

http://www.mersenne.ca/mfaktc.php

there seems to be large differences.

For graphics performance (i.e. Games and stuff), 680 is great. For compute stuff (i.e. CUDA => mmff), not so much. I mean, it is still good, but not great.

henryzz · 2012-11-16, 11:46

Basically 680 is more gaming oriented than the 580. Nvidia have just announced their next version of tesla/server gpus which is computation orientated. Probably the next generation of home cards(7xx) will be based on the same technology. I would guess that 8xx will be gaming mainly etc..

MattcAnderson · 2012-11-17, 16:24

Hi All,
I finally got mmff going on my GPU. The error I had was cudart64_42_9.dll needed. from post # 185 of this thread, I found mmff v 23 with cuda 5.0, which seems to work. This seemed to solve my problem, but the windows 64 download from doublemersenne.org did not. I think this is because it has cuda 4.2. I guess my system requires cuda 5.0 and won't work with cuda 4.2.

Regards,
Matt

2012-11-03, 09:01	#266
aketilander "Åke Tilander" Apr 2011 Sandviken, Sweden 1066₈ Posts	Checksum or residue? The TFs covered by mmff are so large that there will never ever be possible to do a LL to check for compositeness. Therefore I would guess that someone would sooner or later want to do double checks at least when the double mersennes are concerned. Would it be possible to add a checksum or a residue to mmff? Maybe a summarized residue of all the TFs done in a region? I don't think that mmff has a test in the beginning so sometimes, when you don't find factors you start to distrust your system. Did I OC too much? Are there hardware errors? In that case it would be nice to be able to DC a couple of regions just to make sure that your system works OK. So, would it be possible and meaningful to add some kind of checksum or residue to mmff?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Mersenne trial division implementation	mathPuzzles	Math	8	2017-04-21 07:21
trial division over a factor base	Peter Hackman	Factoring	7	2009-10-26 18:27
P95 trial division strategy	SPWorley	Math	8	2009-08-24 23:26
Trial division software for Mersenne	SPWorley	Factoring	7	2009-08-16 00:23
Need GMP trial-division timings	ewmayer	Factoring	7	2008-12-11 22:12

2012-11-03, 14:05	#267
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 11523₈ Posts	I have a question, will mmff client detect the number of GPU's available or not? I mean, should the number of mmff clients be started as many times as the number of GPU cards installed? Thank you in advance, Carlos

2012-11-07, 09:13	#270
akruppa "Nancy" Aug 2002 Alexandria 2,467 Posts	Is there any way to get statistics on the error rate of the trial divison? E.g., what approximate fraction of factors are missed due to race conditions, or due to consumer-grade cards skipping a beat, etc.

2012-11-16, 08:03	#272
aketilander "Åke Tilander" Apr 2011 Sandviken, Sweden 1066₈ Posts	I am wondering about the performance of different videocards. If I compare: http://www.videocardbenchmark.net/high_end_gpus.html with James' http://www.mersenne.ca/mfaktc.php there seems to be large differences. Lets take GTX 680 as an example. According to the first site its on the top of the list, but according to James' its not so good. I can see that James' lists someting called "Compute 3.0". I am not sure if this refers to the speed of the PCI-e bus or if this is something else. What I am really wondering is if there is an unused potential of some videocards which may be better used in the future when mmff is better adapted to "Compute 3.0" or maybe something else?

2012-11-16, 11:46	#274
henryzz Just call me Henry "David" Sep 2007 Cambridge (GMT/BST) 2³×3×5×7² Posts	Basically 680 is more gaming oriented than the 580. Nvidia have just announced their next version of tesla/server gpus which is computation orientated. Probably the next generation of home cards(7xx) will be based on the same technology. I would guess that 8xx will be gaming mainly etc..

2012-11-17, 16:24	#275
MattcAnderson "Matthew Anderson" Dec 2010 Oregon, USA 2⁵×5² Posts	Hi All, I finally got mmff going on my GPU. The error I had was cudart64_42_9.dll needed. from post # 185 of this thread, I found mmff v 23 with cuda 5.0, which seems to work. This seemed to solve my problem, but the windows 64 download from doublemersenne.org did not. I think this is because it has cuda 4.2. I guess my system requires cuda 5.0 and won't work with cuda 4.2. Regards, Matt