Piddling Around With Big Compute Power - Page 2

NBtarheel_33 · 2013-04-26, 21:20

OK, what I have is an 8-core system with no GPU and up to 32 cores and 4 GPUs (2 K20s and 2 K10s) available. 16GB RAM on the former and 32GB on the latter.

Uncwilly, I can P-1 your number if 32GB is enough RAM. Will it need any TF first?

I will also try out CUDA P-1...is it better to DC a known P-1 result or is it ready for production work?

The rep at the NVIDIA dealer providing the trial says that it *may* be possible to arrange future testing on machines with bigger RAM. They have systems with 1TB RAM that retail for $100,000. I think we need three of them...

Uncwilly · 2013-04-27, 03:51

Quote:

Originally Posted by NBtarheel_33

Uncwilly, I can P-1 your number if 32GB is enough RAM. Will it need any TF first?

It is at 79, which is the GPU level for factoring (CPU level is 77). If you bump it up one or two levels, that is fine, either way.

pinhodecarlos · 2013-04-27, 10:13

NBtarheel_33,

Will you be willing to make a benchmark test on your machine with msieve?
Please look at this post

http://www.mersenneforum.org/showpos...ostcount=7%29?

If you are on windows you can substitute the msieve.exe binary with the one in here http://gilchrist.ca/jeff/factoring/; if you are under linux you have it here

http://www.sendspace.com/file/aih4dw

Run under windows

start /low /min msieve.exe -v -nc -t 32

or under linux

./msieve -v -nc -t 32

Thank you in advance,

Carlos

PS( Benchmarks to compare in here http://www.mersenneforum.org/showthread.php?t=16348)

NBtarheel_33 · 2013-04-28, 10:53

So, either I'm dumb or Linux is, but I'm not sure why it is so difficult to spit out how many physical cores are on a system, and whether or not hyperthreading is on, and indeed, whether or not "32 cores available" really just means 16 hyperthreaded physical cores... And so it goes...

I think I have finally figured out the available resources once and for all. I have three systems. The first has 16 non-hyperthreaded cores, 16 GB RAM, and no GPUs. The second has 16 hyperthreaded cores, 32 GB RAM, and two K20 GPUs. And the third has 16 hyperthreaded cores, 32 GB RAM, and two K10 GPUs.

After several days of playing around with settings in mprime and trying to figure out just how in the hell /proc/cpuinfo counts and maps logical vs. physical CPUs, I believe that the best performance for running LLs comes by giving an entire CPU (i.e. 8 cores or 16 hyperthreads) to each LL. In the 50M range, this is yielding iteration times of 4.5ms or so, or 225,000 seconds = 2.6 days for the whole test. 30M doublechecks will complete in just under 24 hours. Suddenly what was a backlog of assignments is turning into not having enough to keep these beasts fed!

On the GPUs - running CUDALucas - I have a 54M and a 56M running on the K20s @ 4.1ms/iteration...which isn't really too much better than an 8-threaded CPU LL test. On the K10s, I am running a 49M @ 9.8ms/iteration, and an 82M @ 14.3ms/iteration. Seems as though if you have big CPU power, it's not as efficient to LL on the GPU...maybe the GPUs are better used for TF or P-1.

By the way, I think I grabbed up an old version of Prime95, because I have run into the dreaded huge roundoff bug. It doesn't actually damage the result, AFAICT, correct?

NBtarheel_33 · 2013-04-28, 11:00

Quote:

Originally Posted by Uncwilly

It is at 79, which is the GPU level for factoring (CPU level is 77). If you bump it up one or two levels, that is fine, either way.

K20s will be free in about 40 hours

. I will take it at least to 80, maybe 81, then try P-1 with 32GB RAM on at least 8 cores, if not 16. (Of course, I have no idea how long P-1 will take, but I have login privileges until May 18th...)

NBtarheel_33 · 2013-04-28, 11:01

Quote:

Originally Posted by pinhodecarlos

NBtarheel_33,

Will you be willing to make a benchmark test on your machine with msieve?
Please look at this post

http://www.mersenneforum.org/showpos...ostcount=7%29?

If you are on windows you can substitute the msieve.exe binary with the one in here http://gilchrist.ca/jeff/factoring/; if you are under linux you have it here

http://www.sendspace.com/file/aih4dw

Run under windows

start /low /min msieve.exe -v -nc -t 32

or under linux

./msieve -v -nc -t 32

Thank you in advance,

Carlos

PS( Benchmarks to compare in here http://www.mersenneforum.org/showthread.php?t=16348)

I'd be happy to...does this just run on CPUs, or GPUs as well?

NBtarheel_33 · 2013-04-28, 11:03

Quote:

Originally Posted by frmky

CUDAPm1 on the K20 would be helpful.

Yes, I'm interested in trying this. I have some P-1's up in the 80-90M range that I'm thinking of trying.

NBtarheel_33 · 2013-04-28, 11:06

Quote:

Originally Posted by Puzzle-Peter

Does it behave like one big machine, i.e. can software be run parallellized over several nodes? If so, try to prove 2^73845+14717 from the Five Or Bust project using PRIMO on 32 cores. It will still take many weeks though.

Unfortunately, there seems to only be the one master login node and two 16-core nodes that I can access. There doesn't seem to be a good way to put the two systems together to get a virtual 32-core system (this would be very nice, but I don't think they are going to go to all that trouble for a test cluster).

NBtarheel_33 · 2013-04-28, 11:09

LOL, I wonder if they realized what kind of "Test Drive" we'd devise to put their systems through. I keep waiting for the "Hmm, you're going to have to scale back your demands on the system" e-mail...

bcp19 · 2013-04-28, 13:08

Quote:

Originally Posted by NBtarheel_33

So, either I'm dumb or Linux is, but I'm not sure why it is so difficult to spit out how many physical cores are on a system, and whether or not hyperthreading is on, and indeed, whether or not "32 cores available" really just means 16 hyperthreaded physical cores... And so it goes...

I think I have finally figured out the available resources once and for all. I have three systems. The first has 16 non-hyperthreaded cores, 16 GB RAM, and no GPUs. The second has 16 hyperthreaded cores, 32 GB RAM, and two K20 GPUs. And the third has 16 hyperthreaded cores, 32 GB RAM, and two K10 GPUs.

After several days of playing around with settings in mprime and trying to figure out just how in the hell /proc/cpuinfo counts and maps logical vs. physical CPUs, I believe that the best performance for running LLs comes by giving an entire CPU (i.e. 8 cores or 16 hyperthreads) to each LL. In the 50M range, this is yielding iteration times of 4.5ms or so, or 225,000 seconds = 2.6 days for the whole test. 30M doublechecks will complete in just under 24 hours. Suddenly what was a backlog of assignments is turning into not having enough to keep these beasts fed!

On the GPUs - running CUDALucas - I have a 54M and a 56M running on the K20s @ 4.1ms/iteration...which isn't really too much better than an 8-threaded CPU LL test. On the K10s, I am running a 49M @ 9.8ms/iteration, and an 82M @ 14.3ms/iteration. Seems as though if you have big CPU power, it's not as efficient to LL on the GPU...maybe the GPUs are better used for TF or P-1.

By the way, I think I grabbed up an old version of Prime95, because I have run into the dreaded huge roundoff bug. It doesn't actually damage the result, AFAICT, correct?

I don't know about mprime that much, but I doubt your best performance is using the entire CPU on a single exponent. Run a benchmark and see the iteration times for 1/2/3/4/5/6/7/8. On mine at 3072k FFT (which is around the 50M exp) I get 16.680 for 1, 8.987 for 2, 6.541 for 3 and 5.842 for all 4. This means with 2 shared cores I only get 92% of those 2 separate, and it drops to 85% with 3 and 71% with all 4.

Prime95 · 2013-04-28, 15:17

Quote:

Originally Posted by NBtarheel_33

By the way, I think I grabbed up an old version of Prime95, because I have run into the dreaded huge roundoff bug. It doesn't actually damage the result, AFAICT, correct?

Fixed in 27.9. All earlier v27s had the bug. You are correct in that it does not damage the result.

2013-04-27, 10:13	#14
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 3×17×97 Posts	NBtarheel_33, Will you be willing to make a benchmark test on your machine with msieve? Please look at this post http://www.mersenneforum.org/showpos...ostcount=7%29? If you are on windows you can substitute the msieve.exe binary with the one in here http://gilchrist.ca/jeff/factoring/; if you are under linux you have it here http://www.sendspace.com/file/aih4dw Run under windows start /low /min msieve.exe -v -nc -t 32 or under linux ./msieve -v -nc -t 32 Thank you in advance, Carlos PS( Benchmarks to compare in here http://www.mersenneforum.org/showthread.php?t=16348) Last fiddled with by pinhodecarlos on 2013-04-27 at 10:17

2013-04-28, 10:53	#15
NBtarheel_33 "Nathan" Jul 2008 Maryland, USA 5×223 Posts	Linux...where 16=32 or is it 16=8? Oh no... So, either I'm dumb or Linux is, but I'm not sure why it is so difficult to spit out how many physical cores are on a system, and whether or not hyperthreading is on, and indeed, whether or not "32 cores available" really just means 16 hyperthreaded physical cores... And so it goes... I think I have finally figured out the available resources once and for all. I have three systems. The first has 16 non-hyperthreaded cores, 16 GB RAM, and no GPUs. The second has 16 hyperthreaded cores, 32 GB RAM, and two K20 GPUs. And the third has 16 hyperthreaded cores, 32 GB RAM, and two K10 GPUs. After several days of playing around with settings in mprime and trying to figure out just how in the hell /proc/cpuinfo counts and maps logical vs. physical CPUs, I believe that the best performance for running LLs comes by giving an entire CPU (i.e. 8 cores or 16 hyperthreads) to each LL. In the 50M range, this is yielding iteration times of 4.5ms or so, or 225,000 seconds = 2.6 days for the whole test. 30M doublechecks will complete in just under 24 hours. Suddenly what was a backlog of assignments is turning into not having enough to keep these beasts fed! On the GPUs - running CUDALucas - I have a 54M and a 56M running on the K20s @ 4.1ms/iteration...which isn't really too much better than an 8-threaded CPU LL test. On the K10s, I am running a 49M @ 9.8ms/iteration, and an 82M @ 14.3ms/iteration. Seems as though if you have big CPU power, it's not as efficient to LL on the GPU...maybe the GPUs are better used for TF or P-1. By the way, I think I grabbed up an old version of Prime95, because I have run into the dreaded huge roundoff bug. It doesn't actually damage the result, AFAICT, correct?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
New GPU Compute System	airsquirrels	GPU Computing	90	2017-12-08 00:13
Analog hardware to compute FFT's...	WraithX	Hardware	1	2012-11-28 13:29
Doubled compute power for a day?	Christenson	PrimeNet	19	2011-10-26 08:29
New Compute Box	Christenson	Hardware	0	2011-01-15 04:44
My throughput does not compute...	petrw1	Hardware	9	2007-08-13 14:38

2013-04-26, 21:20	#12
NBtarheel_33 "Nathan" Jul 2008 Maryland, USA 1115₁₀ Posts	OK, what I have is an 8-core system with no GPU and up to 32 cores and 4 GPUs (2 K20s and 2 K10s) available. 16GB RAM on the former and 32GB on the latter. Uncwilly, I can P-1 your number if 32GB is enough RAM. Will it need any TF first? I will also try out CUDA P-1...is it better to DC a known P-1 result or is it ready for production work? The rep at the NVIDIA dealer providing the trial says that it may be possible to arrange future testing on machines with bigger RAM. They have systems with 1TB RAM that retail for $100,000. I think we need three of them...

2013-04-28, 11:09	#20
NBtarheel_33 "Nathan" Jul 2008 Maryland, USA 1115₁₀ Posts	LOL, I wonder if they realized what kind of "Test Drive" we'd devise to put their systems through. I keep waiting for the "Hmm, you're going to have to scale back your demands on the system" e-mail...