mersenneforum.org  

Go Back   mersenneforum.org > Prime Search Projects > No Prime Left Behind

Reply
 
Thread Tools
Old 2010-08-15, 02:25   #100
vaughan
 
vaughan's Avatar
 
Jan 2005
Sydney, Australia

5×67 Posts
Default

Thanks, that worked.
vaughan is offline   Reply With Quote
Old 2010-09-19, 20:14   #101
Brucifer
 
Brucifer's Avatar
 
Dec 2005

313 Posts
Default

yup, I know, a voice from the past. :-)

So where can I find a current linux llrnet client? How about a linux prpnet client? Are these 32-bit or 64?

I used to bug Gary and Max about the double precision gpu's and wondering if any work had been done on utilizing them for sieving. Has any progress been made on that during the time I've been away from the project?
Brucifer is offline   Reply With Quote
Old 2010-09-19, 23:56   #102
mdettweiler
A Sunny Moo
 
mdettweiler's Avatar
 
Aug 2007
USA (GMT-5)

3×2,083 Posts
Default

Quote:
Originally Posted by Brucifer View Post
yup, I know, a voice from the past. :-)

So where can I find a current linux llrnet client? How about a linux prpnet client? Are these 32-bit or 64?

I used to bug Gary and Max about the double precision gpu's and wondering if any work had been done on utilizing them for sieving. Has any progress been made on that during the time I've been away from the project?
The latest LLRnet client can be found for both Windows and Linux at: http://www.mersenneforum.org/showthread.php?t=13165

Ditto for PRPnet at: http://www.mersenneforum.org/showthread.php?t=12223 (yes, my instructions for that are kind of in a semi-outdated jumble by now...but you're already familiar with the setup so that shouldn't be a problem).

As for GPUs: this time we actually do have some progress to report in that area! Over the last number of months the PrimeGrid guys have developed a sieving program called ppsieve which comes in both CPU and GPU (CUDA, with OpenCL in alpha testing too) versions. Even the CPU version is supposed to be quite a bit faster than sr2sieve for the k-heavy sieves NPLB does, so we're eager to try that our on our next big sieve (k=400-1001, n=1M-2M--planned for the somewhat near future).

As of yet there isn't a GPU application for doing LLR tests, but does exist a program that can do the closely-related LL tests (used by GIMPS) on CUDA GPUs. We're hoping that it can be modified for LLR tests. To that end, Gary has picked up an nVidia GTX 460 GPU on which we can help with testing such an application (and in the meantime, crank out some serious sieving).

Are the GPUs you have ATI or nVidia? Either should be able to help with sieving, though the OpenCL/ATI ppsieve is still in the active development phase and may be a bit of an adventure to work with.

Last fiddled with by mdettweiler on 2010-09-19 at 23:57
mdettweiler is offline   Reply With Quote
Old 2010-09-20, 00:46   #103
Brucifer
 
Brucifer's Avatar
 
Dec 2005

313 Posts
Default

Quote:
Originally Posted by mdettweiler View Post
Are the GPUs you have ATI or nVidia? Either should be able to help with sieving, though the OpenCL/ATI ppsieve is still in the active development phase and may be a bit of an adventure to work with.
I've got both cuda and ati. in the ati, a couple of them are the double precision ones. Are these apps you are testing goint to be windows or linux?
Brucifer is offline   Reply With Quote
Old 2010-09-20, 02:47   #104
mdettweiler
A Sunny Moo
 
mdettweiler's Avatar
 
Aug 2007
USA (GMT-5)

3×2,083 Posts
Default

Quote:
Originally Posted by Brucifer View Post
I've got both cuda and ati. in the ati, a couple of them are the double precision ones. Are these apps you are testing goint to be windows or linux?
Both Windows and Linux are available for ppsieve-CUDA; and I believe both are available for ppsieve-OpenCL as well.

Since this is sieving, the calculations being done are on integers; single/double-precision floating point support is immaterial. It's only for LL/LLR tests and the like that double precision is a must.
mdettweiler is offline   Reply With Quote
Old 2010-09-20, 12:21   #105
vaughan
 
vaughan's Avatar
 
Jan 2005
Sydney, Australia

5×67 Posts
Default

Oooh goodie. GPU crunching. Bring it on.
vaughan is offline   Reply With Quote
Old 2010-09-21, 06:23   #106
Brucifer
 
Brucifer's Avatar
 
Dec 2005

1001110012 Posts
Default

yeah, it would be interesting then cause I've got 9 gpu's that it would be interesting to try doing a sieving run on. They should really speed up the process from what it was like when I was doing a bunch of sieving before on quads. Used to take a ton of time.

So it would be a kick to give stuff a dry run on some already been sieved stuff to see what the comparison is like for speed and accuracy. I've got a trip starting up in another day but will be back next week and it would be a interesting to pursue this if you are game. Just possible that we might be able to really knock out some serious sieving!!!!!

Edit: but I would also like to try out some llr tests if that is at all possible............... heh heh heh

Last fiddled with by Brucifer on 2010-09-21 at 06:24
Brucifer is offline   Reply With Quote
Old 2010-09-21, 07:05   #107
mdettweiler
A Sunny Moo
 
mdettweiler's Avatar
 
Aug 2007
USA (GMT-5)

141518 Posts
Default

Quote:
Originally Posted by Brucifer View Post
yeah, it would be interesting then cause I've got 9 gpu's that it would be interesting to try doing a sieving run on. They should really speed up the process from what it was like when I was doing a bunch of sieving before on quads. Used to take a ton of time.

So it would be a kick to give stuff a dry run on some already been sieved stuff to see what the comparison is like for speed and accuracy. I've got a trip starting up in another day but will be back next week and it would be a interesting to pursue this if you are game. Just possible that we might be able to really knock out some serious sieving!!!!!

Edit: but I would also like to try out some llr tests if that is at all possible............... heh heh heh
Currently NPLB has no active team sieves, but in light of the currently available GPU sieving options, Gary and I have discussed starting our next big sieve, k=400-1001, n=1M-2M, sooner rather than later.

ppsieve-CUDA is already well enough tested that only minimal time needs to be spent on a "dry run" to ensure everything is working properly before jumping into real work. You can download ppsieve-CUDA and find information on how to do a quick test here. If that produces the expected factors then you should be good.

For ppsieve-OpenCL, development has proceeded at least far enough that it can be confidently expected to produce good factors; see here to download it and run a self-test as before with CUDA. The main thing that's left to be done now is to further tweak the program to optimize it better for speed. But it should still be decently fast as is.

With GPUs now so readily usable for sieving, we may actually want to consider using a GPU to calculate the optimal depth for future NPLB sieves. That is, sieve until the removal rate on a GPU is equal to the time it takes to do an equivalent LLR test. We'd be spending just about as much wall-clock time sieving, but we'd have a much higher final depth and a much better sieved final file. Kind of like how we used to calculate optimal depth based on 32-bit sieving, but now use 64-bit.

Last fiddled with by mdettweiler on 2010-09-21 at 07:07
mdettweiler is offline   Reply With Quote
Old 2010-09-21, 07:44   #108
gd_barnes
 
gd_barnes's Avatar
 
May 2007
Kansas; USA

2×52×7×29 Posts
Default

I'm in favor of starting a team sieve on k=400-1001 right away with GPUs if Bruce and Vaughan are available with them. From what Max is indicating, 9 GPUs would do amazing things with such a sieve. Here is what I would do:

1. Sieve k=400-1001 to P=140T.
2. Combine k=300-400 (already sieved to P=140T) from the individual-k drive with k=400-1001.
3. Continue sieving k=300-1001.

That will save some future testing time on k=300-400 also.

Optimal depth might be something ridiculous like P=500T or 750T but it would take us no more GPU hours than it would CPU hours to get to P=140T for the entire range.

All of this begs the question of resources though. Is one GPU like one CPU core? If so, we really only have about 10-15 total GPU cores that we could throw at it at this time and I wouldn't want to monopolize Bruce's and Vaughan's GPU hours so it would be less available "cores" than that.

Last fiddled with by gd_barnes on 2010-09-21 at 07:50
gd_barnes is offline   Reply With Quote
Old 2010-09-21, 15:52   #109
Brucifer
 
Brucifer's Avatar
 
Dec 2005

313 Posts
Default

I'm heading off on a trip with the wifey-poo and daughter :-) leaving tomorrow morning and will be back Sunday evening. So next week I'd like to mess around and see how the cuda thing runs. At the moment I only have two cuda gpu's, having finally burned one up that just worked itself to death. The others are ATI cards. As I mentioned in my PM to Gary, I went reading the thread at PG and it was interesting. I do get the impression though that there is still some speed optimization issues, but then too it's very possible that the approach being taken, or the whole basic issue doesn't lend itself readily to parallel crunching... don't know cause I haven't messed with it. So one part did mention though that one was still faster than a quad so that's still a gain. But yup, I'm pretty hyped up about it after doing all that prior sieving the long slow hard way, this will be something new the experience. :-) Much of that testing was done on not-so-new cuda cards so those were still ripping along pretty good, considering. Then the new card like Gary's, I have no idea how those run as I haven't invested in one of them yet. :-)

edit: I see that the url above Max goes to the PG thread. The two linux ppsieve(cuda) at the top are the early initial testing source for linux. Farther down in the thread they are talking about XP. Is there any precompiled binaries for XP for the cuda card, and if so, which cuda release is it running under... 3.1??

Last fiddled with by Brucifer on 2010-09-21 at 16:00
Brucifer is offline   Reply With Quote
Old 2010-09-21, 18:46   #110
mdettweiler
A Sunny Moo
 
mdettweiler's Avatar
 
Aug 2007
USA (GMT-5)

3·2,083 Posts
Default

Quote:
Originally Posted by gd_barnes View Post
All of this begs the question of resources though. Is one GPU like one CPU core? If so, we really only have about 10-15 total GPU cores that we could throw at it at this time and I wouldn't want to monopolize Bruce's and Vaughan's GPU hours so it would be less available "cores" than that.
Short answer: yes. One GPU core behaves like a single unit.

Long answer: sometimes. Depending on the program you're running and how well optimized it is, sometimes you can actually get more throughput if you run more than one instance of it on the same GPU. However, I do not believe this is the case with ppsieve, since in my testing it is mostly GPU bound (that is, the slowest link is the GPU; not the small prime sieve portion of the algorithm which is being run on the CPU. For tpsieve, the situation is reversed and the application is CPU bound, taking almost an entire CPU core in addition to the GPU when running with properly optimized settings.)

So as far as ppsieve goes, yes, it would monopolize the GPU on which it is running.
mdettweiler is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Some old stuff Batalov Miscellaneous Math 1 2017-01-27 04:56
Stuff for sale fivemack Lounge 12 2011-06-12 11:28
useful stuff paulunderwood Linux 3 2005-12-05 22:18
Free stuff... Xyzzy Software 6 2004-10-06 13:35
Extra Stuff... Xyzzy Lounge 11 2003-09-15 23:22

All times are UTC. The time now is 21:53.

Thu Jul 9 21:53:41 UTC 2020 up 106 days, 19:26, 0 users, load averages: 0.89, 1.30, 1.49

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.