E@h and S@h simultaneously on one GPU?
Message boards :
Number crunching :
E@h and S@h simultaneously on one GPU?
Message board moderation
| Author | Message |
|---|---|
 Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5 
|
Hello community! I know someone, he let run Einstein@home on a GTX260-192. He see only ~ 9 % GPU usage of the CUDA app there. If he crunch SETI@home he see ~ 80 % usage. We are curious about, if it's possible to crunch 2x E@h and 1x S@h WUs on this/one GPU. It's possible? Performance loss? How should look the app_info.xml for E@h? It's correct, that as E@h member you must crunch everything what's possible/available? You can't choose which app you want crunch? How should look the other settings that then are 3 (2x E@h + 1x S@h) WUs on one GPU? Thanks! 
ID: 1028151 ·  |
|
Gundolf Jahn Send message Joined: 19 Sep 00 Posts: 3184 Credit: 446,358 RAC: 0
|
Is it correct, that as E@h member you must crunch everything what's possible/available? You can't choose which app you want crunch? There are plenty of threads on that topic at the Einstein boards! You can choose if you want to run ABPS or not, since it's an addon. You can't deselect S5GC in the preferences, since that's the application for which Einstein@home gets funded, but of course there are other possibilities. Gruß, Gundolf Computer sind nicht alles im Leben. (Kleiner Scherz)  SETI@home classic workunits 3,758 SETI@home classic CPU time 66,520 hours ID: 1028229 · |
|
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5
|
My english is very poor. He's also a non native english speaker. So it's not easy to find certain infos. Because of I have no E@h account, I thought the friendly S@h members could help. Which run also E@h. He worried about, that if we make an app_info.xml file and the admins see (however) that he make cherry-picking with apps that he could get troubles. I don't know E@h. So I asked about this. Noone who crunch E@h on a nVIDIA GPU here? It's well known that the E@h CUDA app is currently not big optimized. But to now, noone made (published) an app_info.xml file for 'newcomers'. We don't know how to make an app_info.xml file for E@h. Sure, I can do it for S@h.. no problem.. But I don't know the project name of E@h for the file also not the correct names/version numbers of the apps. We don't want to trash WUs, so we are smart and ask before we do something 'bad'. And the question is still, if it's possible to crunch 2 different projects simultaneoulsy on one GPU. ;-) Thanks! EDIT: BTW. O.K., I know the apps: http://einstein.phys.uwm.edu/apps.php But for to exclude mistakes, it would be well to ask first people with knowledge, right? ;-)
ID: 1028284 · |
|
ront Send message Joined: 25 Aug 01 Posts: 77 Credit: 386,336 RAC: 0 
|
Good Morning, Was working an "Astropulse" WU with about 60 hours left to process. Woke up this morning to see that processing had stopped at 108 hours and some minutes. The notation was "computational error." ? Anybody able to explain what this means? Thanks for the help & assistence. ron ID: 1028294 · |
|
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5
|
Hmm? Sorry, you know the forum..
* Posts must be 'kid friendly': they may not contain content that is obscene, hate-related, sexually explicit or suggestive. * No commercial advertisements. * No links to web sites involving sexual content, gambling, or intolerance of others. * No messages intended to annoy or antagonize other people, or to hijack a thread. * No messages that are deliberately hostile or insulting. * No abusive comments involving race, religion, nationality, gender, class or sexuality. * Do not post pictures of graphic violence or links to images or videos containing graphic violence. * Do not posts comments related to specific moderator actions. The moderators may be contacted at seti_moderators(X)ssl(dot)berkeley(dot)edu * Your posts may be deleted if it contains material from another post which has been deleted. You should avoid including material from a post you expect will be deleted. * Team recruiting is only allowed in the Team Recruitment Forum * Additional information on moderation policy can be found here.
ID: 1028301 · |
Jord  Send message Joined: 9 Jun 99 Posts: 15184 Credit: 4,362,181 RAC: 3 
|
It's well known that the E@h CUDA app is currently not big optimized. Optimized as in using the CPU instructions (SSE, SSE2 etc.)? Both GC and ABP2 are in that sense optimized. GC will use either SSE, SSE2 or none of the before depending on what CPU instructions have been detected and sent to the server. It will then send the correct science application your way. ABP2 requires an SSE capable CPU. Without it you will not get ABP2 work. ABP2 CUDA will only use the GPU for the FFT algorithm calculations, the rest of the calculations are still done on the CPU. The FFT algorithm will use up about 300MB of Video RAM, hence the requirement of CUDA cards with 512MB or more memory. ID: 1028388 · |
|
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5
|
Ahh.. O.K., thanks! So the E@h CUDA app is like the AP CPU/GPU hybrid ATI app? So not really a pure GPU app. I read, that E@h CUDA WUs are shown: (1.00 CPUs + 1.00 GPUs) in BOINC. So we need to reduce the GPUs value. AFAIK, this is possbile over the app_info.xml coproc value: <coproc> <type>CUDA</type> <count>1</count> </coproc> He see ~ 9 % GPU usage.. then maybe: <coproc> <type>CUDA</type> <count>0.1</count> </coproc> ..for 10 % on the GPU. But maybe now BOINC will calculate 10 E@h WUs on the CPU? He have a Quad-CPU.. not the new i7 12 thread CPU.. ;-) Ohh.. I think it's not possible to crunch E@h and S@h simultaneously on one GPU. Because I see no way to adjust this. I guess the max would be to crunch 4x E@h on CPU and 'outsource' calculation to the one GPU. So 4x E@h CUDA WUs. Or? Thanks!
ID: 1028439 · |
JohnDK    Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127 
 |
I tried running S@H + E@H on one GPU. Got the app info from client_state.xml and made a app_info.xml looking like this <app_info> Also changed CUDA count to 0.9 in SETI app_info. Not sure if the WU runs at same speed as they did before individually, but think it looks good... There's a BIG but... When I downloaded the einstein WU, I got this message Message from server: To get more Einstein@Home work, finish current work, stop BOINC, remove app_info.xml file, and restart. Right now I'm running both cuda WUs but if Einstein will get a new WU when this is finished I don't know... ID: 1028458 · |
Fred J. Verster  Send message Joined: 21 Apr 04 Posts: 3252 Credit: 31,903,643 RAC: 0
|
I've Einstein as one of my 'Back Up Projects and just have an GTX470 installed, appeared to be 'incompatible with an 9800GTX+ .(WIN XP64; QX9650 CPU+GTX470) Einstein CUDA uses the GPU, but little is done and also takes a full core. Gonna try that too, trying to run more then 1 WU at a time, on a FERMI, it is possible. And now I can test new optimized MB for FERMI and ATI for AstroPulse....... @ Sutaru , why not open an account at Einstein ;^)........ [ADDED]Einstein doesn't accept work done by 'modified application', they provide the work and application it's a policy.......... 
ID: 1028464 · |
|
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5
|
Thanks! Ohh.. maybe E@h don't want an app_info.xml file in the project folder?
ID: 1028470 · |
|
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5
|
Now I see your EDIT. ;-) Really, E@h don't want an app_info.xml file in the project folder? If, then it's not possible to adjust 2+ WUs on one GPU at E@h. Or 1x S@h and 1x E@h on one GPU - I guess. Because if the BOINC Manager make (1.00 CPUs + 1.00 GPUs) for E@h CUDA the whole GPU is used and no space for more, or? Maybe someone have values for to compare if the E@h CUDA hybrid app is faster than the pure E@h CPU app? Which CPU and which GPU? Thanks!
ID: 1028471 · |
|
Fred J. Verster Send message Joined: 21 Apr 04 Posts: 3252 Credit: 31,903,643 RAC: 0
|
Look at the posts above ;-) That's true, 1 WU takes 1 GPU + 1 CPU (core). That is a consequence of that policy, no app_info.xml file needed to run stock app's. I find it a bit conservative, why not allow an optimized app. if it produces valid result's? But that also takes extra time, which isn't available, so nothing changes, as a result, IMO :)
ID: 1028474 · |
|
JohnDK Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127
|
There's a tread on Einstien where they talk about app_info and how it's "cheating", since some apps gets more credits than others... ID: 1028479 · |
|
JohnDK Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127
|
Was on my way to bed when I thought, why not just try changing the CUDA count in client_state.xml instead of using the app_info. So far it works, running Seti & Einstein on one GPU. It remains to be seen if the CUDA count stay put when I get new WUs or if it resets itself... ID: 1028510 · |
|
garfield Send message Joined: 4 Jan 02 Posts: 45 Credit: 7,409,265 RAC: 65 
 |
Was on my way to bed when I thought, why not just try changing the CUDA count in client_state.xml instead of using the app_info. So far it works, running Seti & Einstein on one GPU. It remains to be seen if the CUDA count stay put when I get new WUs or if it resets itself... Hi, could you please post where exact you made the changes? There are so many lines of einstein-entries! I have some seti cuda wu's left from tuesday, I could check that today! Regards, Alexander ID: 1028634 · |
|
JohnDK Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127
|
Hi, It's in the section with the einstein cuda entries. <app_version> ID: 1028664 · |
|
garfield Send message Joined: 4 Jan 02 Posts: 45 Credit: 7,409,265 RAC: 65
|
Hi, many, many thanks! There is one additional info: Einstein places two CUDA-apps in the client_info.xml. Both <count>'s have to be changed, otherwise it is immediate reset to 1.000 after starting. At the moment I can run one seti-cuda 0.04CPU / 0.5 GPU and two Einstein 1 CPU / 0.25 GPU. Tomorrow, when seti is online again, I will see whether or not the results of seti are correct; info about Einstein will be available earlier. And many thanks to Sutaru as well for his effort to get these Infos. Regards, Alexander ID: 1028682 · |
|
JohnDK Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127
|
There is one additional info: Einstein places two CUDA-apps in the client_info.xml. Both <count>'s have to be changed, otherwise it is immediate reset to 1.000 after starting. I only have one Einstein cuda count, so maybe you have two different Einstein cuda apps running. Mine didn't reset after starting the WU/Boinc. ID: 1028683 · |
|
JohnDK Send message Joined: 28 May 00 Posts: 1222 Credit: 451,243,443 RAC: 1,127
|
At the moment I can run one seti-cuda 0.04CPU / 0.5 GPU and two Einstein 1 CPU / 0.25 GPU. Apparently Einstein doesn't use that much GPU, so I guess 0.25 is to much, it means seti cuda runs slower than it can. Sutaru said Einstein only uses about 9% in the first post. ID: 1028684 · |
|
garfield Send message Joined: 4 Jan 02 Posts: 45 Credit: 7,409,265 RAC: 65
|
You're right. Currently I'm experimenting with the settings. Since Einstein uses also a complete CPU it is a target for the task-switcher and so sometimes removed, causing two seti's to run simultanously. Afterburner shows a lower GPU-Usage when running two seti's together, so I changed the setting now to 0.55. Running one single seti-cuda Afterburner shows the same 90% usage even when set to 0.55. This leaves the possibility to run always only one single seti-wu and, if the taskswitcher wants that, up to two additional Einsteins (now set to 0.22 GPU's). Due to memory usage it might make no sense to run more than three apps at a time. One problem leaves: When Boinc is stopped and restarted, the <count> is reset to 1. Alexander ID: 1028691 · |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.