mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware

Reply
 
Thread Tools
Old 2005-11-24, 22:14   #1
ixfd64
Bemusing Prompter
 
ixfd64's Avatar
 
"Danny"
Dec 2002
California

23·33·11 Posts
Default dumb question - getting multiple processors to work together?

I know that people are saying that the L-L test is more suitable for a single processor than a supercomputer, etc. The L-L test uses the result from the previous step, so using many parallel processors would not be so practical.

However, on the GIMPS main page, it says that a Bull NovaScale computer with multiple processors is used successfully.

Does this have anything to do with the code and/or the structure of the processors? I am very interested in this.

Thanks.

Last fiddled with by ixfd64 on 2005-11-24 at 22:22
ixfd64 is offline   Reply With Quote
Old 2005-11-24, 22:43   #2
Peter Nelson
 
Peter Nelson's Avatar
 
Oct 2004

10000100012 Posts
Default

Each LL iteration must occur after the previous is completed.

However there is much calculation in each iteration which is done by a complex FFT.

It is possible to spread this work across multiple cpus (or parts of a chip which run in parallel like doing the fft on fpga chips).

However the work of sharing out the math between processors involves inter-processor communications, which are not needed on a single processor. Therefore performance does not scale linearly with number of processors.

The rationale for prime95/mprime working on one exponent on one processor is that the other processors could be running separate instances of prime95 on different exponents.

Therefore we avoid the performance hit of the sharing, and although exponents still take a LONG time, in terms of the contribution to the project overall it is higher.

One time when it is useful to sacrifice performance per cpu for overall speed to complete is when verifying a suspected mersenne prime candidate.

That is why multiple-cpu machines are used for that purpose.

Personally I think that as we test bigger exponents which take longer per iteration and longer overall, then there will come a size when it takes longer to complete an LL that is desirable because people will get bored waiting and/or the machine may get retired and the work lost.

Therefore I consider it desirable if the current prime95/mprime were to optionally support to run in a multi-cpu-cooperating mode.
Peter Nelson is offline   Reply With Quote
Old 2005-11-25, 01:31   #3
drew
 
drew's Avatar
 
Jun 2005

2×191 Posts
Default

Quote:
Originally Posted by Peter Nelson
Each LL iteration must occur after the previous is completed.

However there is much calculation in each iteration which is done by a complex FFT.

It is possible to spread this work across multiple cpus (or parts of a chip which run in parallel like doing the fft on fpga chips).

However the work of sharing out the math between processors involves inter-processor communications, which are not needed on a single processor. Therefore performance does not scale linearly with number of processors.

The rationale for prime95/mprime working on one exponent on one processor is that the other processors could be running separate instances of prime95 on different exponents.

Therefore we avoid the performance hit of the sharing, and although exponents still take a LONG time, in terms of the contribution to the project overall it is higher.

One time when it is useful to sacrifice performance per cpu for overall speed to complete is when verifying a suspected mersenne prime candidate.

That is why multiple-cpu machines are used for that purpose.

Personally I think that as we test bigger exponents which take longer per iteration and longer overall, then there will come a size when it takes longer to complete an LL that is desirable because people will get bored waiting and/or the machine may get retired and the work lost.

Therefore I consider it desirable if the current prime95/mprime were to optionally support to run in a multi-cpu-cooperating mode.
The current client and server software doesn't have the capability to transfer tests from one PC to another, does it? That may become necessary for the reason you described. I would be nice if during check-in, the current iteration and residue get stored so that if it times-out, someone else can continue.

Drew
drew is offline   Reply With Quote
Old 2005-11-25, 23:24   #4
Peter Nelson
 
Peter Nelson's Avatar
 
Oct 2004

21116 Posts
Default

Quote:
Originally Posted by drew
The current client and server software doesn't have the capability to transfer tests from one PC to another, does it? That may become necessary for the reason you described. I would be nice if during check-in, the current iteration and residue get stored so that if it times-out, someone else can continue.

Drew
I THINK that do save and continue on a different machine may need SEVERAL MEGABYTES of save data. I think this is the reason this feature is not supported so that the server can handle the net traffic (or storage space) from all clients doing this.

This would be impractical to do routinely, BUT might be acceptable if a manual option were added to send the save file when a machine retires from the project.

Of course there would then be the problem of identifying which of (2+) machines may have produced an error where results don't match a double check.

I still think multi-cpu prime95 may be the best approach.
Peter Nelson is offline   Reply With Quote
Old 2005-11-26, 00:33   #5
drew
 
drew's Avatar
 
Jun 2005

2·191 Posts
Default

Quote:
Originally Posted by Peter Nelson
I THINK that do save and continue on a different machine may need SEVERAL MEGABYTES of save data. I think this is the reason this feature is not supported so that the server can handle the net traffic (or storage space) from all clients doing this.
Good point. Since these values are mod(n), they should be the same size as the Mersenne number being tested, which is huge. That's a lot of storage space and bandwidth I didn't consider.

Drew

Last fiddled with by drew on 2005-11-26 at 00:41
drew is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
N00B Question about system with multiple processors QuickCoder Hardware 15 2014-08-10 18:45
Is this a dumb question? davieddy Science & Technology 9 2012-02-23 10:53
Ask a dumb question... davieddy Information & Answers 17 2011-08-06 13:34
Which versions for multiple processors, Xeon? kdq Information & Answers 5 2008-09-28 06:18
Prime95, hyperthreading, multiple processors, Win2003, etc... pcr Software 8 2005-12-22 14:43

All times are UTC. The time now is 00:52.

Mon Apr 12 00:52:18 UTC 2021 up 3 days, 19:33, 1 user, load averages: 1.30, 1.40, 1.46

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, 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.