New Monster in the making.

Message boards : GPUs : New Monster in the making.
Message board moderation

To post messages, you must log in.

AuthorMessage
jesse

Send message
Joined: 11 Jun 14
Posts: 1
United States
Message 54420 - Posted: 11 Jun 2014, 0:29:55 UTC

Hello all,

I am thinking of constructing a monster/micro super computer. I estimate 22 Tflops roughly as my entry machine.

2x16 core opteron 2.2 ghz socket g34
16GB ram
2x r9 295x2

or maybe near future construct this monster

This system also yeilds double the previous machine at about 44 Tflops roughly.

1x16 core opteron 2.2 ghz socket g34
16GB ram
4x r9 295x2 cross fire

I was wondering are these components capable with boinc?

I will be running windows 7 ultimate leaving this machine to help out SETI in the near future possibly make another one.

I have access to 100 gigbit fiber line. Just wanna help out and have fun building monster machines and donate clock cycles to SETI or what needs it.

Thanks guys.
ID: 54420 · Report as offensive
Coleslaw
Avatar

Send message
Joined: 23 Feb 12
Posts: 182
United States
Message 54473 - Posted: 13 Jun 2014, 20:12:48 UTC - in response to Message 54420.  
Last modified: 13 Jun 2014, 20:17:13 UTC

I don't see why those components wouldn't be compatible with BOINC. I will say that the crossfire setup doesn't help you any with BOINC as far as I know. So, if you are building this as a dedicated crunching box, you may want those cards to remain separate. For that many cores, you will also be light on RAM for some projects, so you may want to consider increasing it to minimum of 1GB/core. Some projects want 3GB+ per work unit. So, I would also recommend a solid state hard drive or else you may burden the system down with too many IO's. You didn't say what storage you were using. Just some food for thought.

I will also suggest running the 64bit version of BOINC because we have seen where the 32bit version doesn't care for anything over 40 cores... Or at least I think it was the 32bit version only. I have not done a lot of testing myself with it. Hopefully soon I will be able to play with that many or more cores to find out exactly. I know that we had to use a cc_config file to tell BOINC the number of cores on one of my team mates rigs recently running Linux.
ID: 54473 · Report as offensive
noderaser
Avatar

Send message
Joined: 2 Jan 14
Posts: 267
United States
Message 54478 - Posted: 14 Jun 2014, 2:06:29 UTC

With the number of cores found in today's crunching machines going up and up, hopefully more projects will adopt multi-threaded applications, those can really be a memory saver.

ID: 54478 · Report as offensive
Coleslaw
Avatar

Send message
Joined: 23 Feb 12
Posts: 182
United States
Message 54556 - Posted: 23 Jun 2014, 16:24:04 UTC - in response to Message 54478.  

With the number of cores found in today's crunching machines going up and up, hopefully more projects will adopt multi-threaded applications, those can really be a memory saver.


They can also create other problems. Yafu is a good example. We (my team) are finding that YAFU's scoring system (Credit New) doesn't reward larger core counts very well. Apparently, a 12 core machine can make almost as many points as a 64core machine. This is terrible. And there are a lot of people who will drop a project based on points alone. I don't know if it is a YAFU issue or a Credit New issue or even a BOINC problem in general. Just an observation we have had since my team has a lot of high end server grade equipment transitioning off of FAH as of late.
ID: 54556 · Report as offensive
Richard Haselgrove
Volunteer moderator
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 2611
United Kingdom
Message 54558 - Posted: 23 Jun 2014, 17:23:46 UTC - in response to Message 54556.  

With the number of cores found in today's crunching machines going up and up, hopefully more projects will adopt multi-threaded applications, those can really be a memory saver.

They can also create other problems. Yafu is a good example. We (my team) are finding that YAFU's scoring system (Credit New) doesn't reward larger core counts very well. Apparently, a 12 core machine can make almost as many points as a 64core machine. This is terrible. And there are a lot of people who will drop a project based on points alone. I don't know if it is a YAFU issue or a Credit New issue or even a BOINC problem in general. Just an observation we have had since my team has a lot of high end server grade equipment transitioning off of FAH as of late.

I don't know the YAFU project, but I have done MT-app testing at both AQUA (now ceased operating) and MilkyWay. If YAFU are using straight CreditNew code (i.e. BOINC server source code, without local modification), then I'd say it's 99% sure that CreditNew is to blame.

There are intensive studies of CreditNew under way at the moment on the 'Albert' project (Einstein@Home's testbed) - see the News thread Project server code update. At the moment we're looking at CreditNew for GPU apps: MT should be studied too, but unfortunately MT is about the only class of apps which Einstein/Albert can't provide a testing environment for. Does YAFU have a beta or similar test environment where server patches could be tried out in a few weeks' time, under observation in a controlled environment?
ID: 54558 · Report as offensive
Coleslaw
Avatar

Send message
Joined: 23 Feb 12
Posts: 182
United States
Message 54559 - Posted: 23 Jun 2014, 17:28:50 UTC - in response to Message 54558.  

With the number of cores found in today's crunching machines going up and up, hopefully more projects will adopt multi-threaded applications, those can really be a memory saver.

They can also create other problems. Yafu is a good example. We (my team) are finding that YAFU's scoring system (Credit New) doesn't reward larger core counts very well. Apparently, a 12 core machine can make almost as many points as a 64core machine. This is terrible. And there are a lot of people who will drop a project based on points alone. I don't know if it is a YAFU issue or a Credit New issue or even a BOINC problem in general. Just an observation we have had since my team has a lot of high end server grade equipment transitioning off of FAH as of late.

I don't know the YAFU project, but I have done MT-app testing at both AQUA (now ceased operating) and MilkyWay. If YAFU are using straight CreditNew code (i.e. BOINC server source code, without local modification), then I'd say it's 99% sure that CreditNew is to blame.

There are intensive studies of CreditNew under way at the moment on the 'Albert' project (Einstein@Home's testbed) - see the News thread Project server code update. At the moment we're looking at CreditNew for GPU apps: MT should be studied too, but unfortunately MT is about the only class of apps which Einstein/Albert can't provide a testing environment for. Does YAFU have a beta or similar test environment where server patches could be tried out in a few weeks' time, under observation in a controlled environment?


YAFU is a alpha project, which main goal is

to test the latest Boinc server code for bugs, please report any error which you find in the webpages in the forum
to factorize numbers of 80-110 digit length which needs to be factored
to test the latest versions of yafu.


YAFU is the test bed.
ID: 54559 · Report as offensive
Richard Haselgrove
Volunteer moderator
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 2611
United Kingdom
Message 54560 - Posted: 23 Jun 2014, 17:32:59 UTC - in response to Message 54559.  

YAFU is the test bed.

Excellent. I'll pass that back to the team, and we may have some patches for you to test. Who should we get in touch with?
ID: 54560 · Report as offensive
Coleslaw
Avatar

Send message
Joined: 23 Feb 12
Posts: 182
United States
Message 54563 - Posted: 23 Jun 2014, 17:49:28 UTC - in response to Message 54560.  

As far as the project goes I believe YoYo. Here is a thread where my team mate mentions benchmarking concerns. He has a lot of high end servers in his home. YoYo replies in that thread.

http://yafu.myfirewall.org/yafu/forum_thread.php?id=160
ID: 54563 · Report as offensive
Richard Haselgrove
Volunteer moderator
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 2611
United Kingdom
Message 54564 - Posted: 23 Jun 2014, 17:55:17 UTC - in response to Message 54563.  

As far as the project goes I believe YoYo. Here is a thread where my team mate mentions benchmarking concerns. He has a lot of high end servers in his home. YoYo replies in that thread.

http://yafu.myfirewall.org/yafu/forum_thread.php?id=160

I've emailed my team about it, but it's a cross-continent and hence cross-timezone effort, and things move slowly.

Another useful thread would be http://yafu.myfirewall.org/yafu/forum_thread.php?id=147
ID: 54564 · Report as offensive

Message boards : GPUs : New Monster in the making.

Copyright © 2018 University of California. 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.