mersenneforum.org mfaktc (0.20)
 Register FAQ Search Today's Posts Mark Forums Read

2013-09-16, 09:26   #1
fairsky

Sep 2013

41 Posts
mfaktc (0.20)

Hi,
Quote:
 ./mfaktc.exe: /lib64/libc.so.6: version GLIBC_2.14' not found (required by ./mfaktc.exe)
Then I think I need to compile it from source. But anyone compiled mfaktc 0.20 successfully?

I tried to compile it, but I also got wrong:
Quote:

Also, my another question on multi-gpu support!
I have 3 gpu units (in 2 cards), how can I run one single job on 3 gpus (this is a cross gpu work!!!!, so dont tell me the readme file or Q/A has given the answer. If I understood this file correctly, it ONLY tells us that each gpu works for a single job, NOT my concern here).

2013-09-16, 09:47   #2
TheJudger

"Oliver"
Mar 2005
Germany

2×3×5×37 Posts

Quote:
 Originally Posted by fairsky Then I think I need to compile it from source. But anyone compiled mfaktc 0.20 successfully?
Yes, I did!

Quote:
 Originally Posted by fairsky I tried to compile it, but I also got wrong: Code: [...] read_config.c:22:26: error: cuda_runtime.h: No such file or directory [...]`
CUDA toolkit installed (and added to environment)?

Quote:
 Originally Posted by fairsky Also, my another question on multi-gpu support! I have 3 gpu units (in 2 cards), how can I run one single job on 3 gpus (this is a cross gpu work!!!!, so dont tell me the readme file or Q/A has given the answer. If I understood this file correctly, it ONLY tells us that each gpu works for a single job, NOT my concern here).
You'll need to patch the code (add multi GPU support), currently this is NOT POSSIBLE (as written in README.txt). And there are no plans to add support for this.

Oliver

Last fiddled with by TheJudger on 2013-09-16 at 09:48

2013-09-16, 10:00   #3
fairsky

Sep 2013

41 Posts

Quote:
 Originally Posted by TheJudger Yes, I did! CUDA toolkit installed (and added to environment)?
Oh, thank you. I think had blind eyes..
Well, I will double check the file you pointed out.

Quote:
 You'll need to patch the code (add multi GPU support), currently this is NOT POSSIBLE (as written in README.txt). And there are no plans to add support for this. Oliver
well, i just read the doc from nvidia, but I am not sure how can i patch it.
(i just noticed the stream mechanism there....)

 2013-09-16, 10:13 #4 fairsky   Sep 2013 41 Posts wow...I checked the Makefile. Then user who wants to compile it should modify the dir for cuda installed. So, now it is done with cuda 4.2.9. thanks oliver for pointing out my mistake.
2013-09-16, 10:42   #5
fairsky

Sep 2013

41 Posts

Quote:
 Originally Posted by fairsky wow...I checked the Makefile. Then user who wants to compile it should modify the dir for cuda installed. So, now it is done with cuda 4.2.9. thanks oliver for pointing out my mistake.
Ohhhh my dear.....
I should have started using mfaktc month before.
I found now mfaktc reduces time largely.
mprime will complete a 320M TF job within one (and more) month with one cpu/thread, about 16days within 8threads, but mfaktc will do it within just 2hrs10mins......

(sorry, my job is 32x,xxx,xxx 75,76).

Last fiddled with by fairsky on 2013-09-16 at 10:57

2013-09-16, 16:27   #6
TheJudger

"Oliver"
Mar 2005
Germany

21268 Posts

Quote:
 Originally Posted by fairsky well, i just read the doc from nvidia, but I am not sure how can i patch it. (i just noticed the stream mechanism there....)
Take this more as a joke. Won't be an trivial task (but it is possible, I just don't see the need to do so).

Quote:
 Originally Posted by fairsky Ohhhh my dear..... I should have started using mfaktc month before. I found now mfaktc reduces time largely. mprime will complete a 320M TF job within one (and more) month with one cpu/thread, about 16days within 8threads, but mfaktc will do it within just 2hrs10mins...... (sorry, my job is 32x,xxx,xxx 75,76).
Yes, currently TF on CPU is obsolete in most cases.

Oliver

2013-09-17, 04:40   #7
TheMawn

May 2013
East. Always East.

11×157 Posts

Compared to the hassle of adding multiple GPU support, it is so ridiculously easy to set up multiple instances of the program. SLI or CFX would only be good for trying to do ONE exponent two or three or four times faster, but they're so fast anyway for probably 99.9999% of the work we actually do that there's no reason for it. There is interest in maybe three or four specific exponents that are having massive TF work done on them, but for the rest, nobody cares if they have to get three exponents after three weeks versus one after one week.

I use a simple one-line batch file.

Quote:
 start E:\...\mfaktc.exe -d 0
You can add some bells and whistles for more functionality, like the speed controller batch file http://mersenneforum.org/misfit/downloads/other/

I added -d 1 to the line

Quote:
 set MF=MFAKTC.EXE -d 1
And that's all I need. In these cases, the 0 and 1 refer to the GPU slot values, so you will likely use 0 1 and 2 for a three-way GPU setup.

You'll have to ask someone here with actual SLI experience to hash out the finer details (i.e. does this still work if the GPU's are bridged and already set up in SLI or CFX?)

 2013-09-17, 04:51 #8 kladner     "Kieren" Jul 2011 In My Own Galaxy! 2·5,059 Posts AFAIK, SLI is not supported in any case. All the talk I remember centered around disabling it, even on GTX x90 boards. Let LaurV weigh in on this to be sure, but the chips have to be addressed separately. Last fiddled with by kladner on 2013-09-17 at 04:53
 2013-09-17, 12:56 #9 LaurV Romulan Interpreter     Jun 2011 Thailand 8,963 Posts Right, if you don't game, and I mean big game, like crisis or so, then SLI is completely unuseful, and it actually makes things slower (communication between cards over the SLI cable). Some "folders" claim benefits from SLI, but honestly, I do better for folding if I try two proteins in the same time with separate cards (I didn't do that for ages, there might be changes, but at the time I tried, I did not see any, I mean ANY, benefits of SLI. Even AutoCad, it is so fast with a single 580, that you won't need two, and other design programs like Altium Protel, OrCad, SolidWorks, they can't take fully advantage of SLI yet (bad/old drivers, AutoDesk is the most advanced in this direction). For mfaktc, especially, there are no advantages. If you use 0.20, with GPU sieving, then one instance per card is enough to max it. If you use older versions, or use CPU sieving, then you may need more instances per card, for example if you have 3 cores free on your CPU, then you launch 3 instances of mfaktc with CPU sieving, and you will get some more speed, but at the penalty that the CPU cores are locked there, i.e. you can't use them to do P95 LL or P-1 work. I like mfaktc with GPU sieving, one instance per card, it has the merit that lets the CPU free for other tasks, and being only a single instance per card, does not require much maintenance (as opposite to older version of mfaktc, where you had to feed 8 or 12 instances, in different folders, etc). You guys, the newcomers, are spoiled with the new mfaktc, MISFIT, and other utilities which make your life easier. (and I use them too, and I am happy with them, but years ago, the life was much harder, hehe). Last fiddled with by LaurV on 2013-09-17 at 12:59 Reason: /s/mfactk/mfaktc
 2013-09-24, 12:58 #10 Karl M Johnson     Mar 2010 1100110112 Posts Multi-gpu support may be required when having multiple gpus becomes mainstream. That is not going to happen within a couple of years, as transistor miniaturization requires more and more investment. Also, trial factoring a single exponent at a very high bit level may benefit from a bunch of gpus (but that needs to be justified versus a LL test). So, fairsky, as you can see, a script that distributes the workload across N mfaktc instances is a an easier and a faster solution.

 Similar Threads Thread Thread Starter Forum Replies Last Post Peter GPU Computing 19 2014-10-19 15:26 bayanne GPU Computing 0 2014-05-10 14:38 bayanne GPU Computing 0 2013-10-18 09:59 NormanRKN GPU Computing 24 2012-07-29 15:25 tichy GPU Computing 4 2010-12-03 21:51

All times are UTC. The time now is 03:19.

Sat Dec 5 03:19:20 UTC 2020 up 1 day, 23:30, 0 users, load averages: 1.83, 1.77, 1.64