| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: OpenPandemics - COVID-19 Project
Thread: OpenPandemics - GPU Stress Test |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 781
|
|
| Author |
|
|
Michael Goetz
Cruncher United States Joined: Dec 11, 2017 Post Count: 35 Status: Offline Project Badges: 
|
The 13510 series keeps 1 core of my AMD 1700 Ryzen busy full time. My NVidia GTX 1060 only run sometimes. So CPU limited. 1 case now takes 14 minutes (older cases about 4-5 minutes). nvidia openCL applications always keep a core at 100%. This is no longer true. While *many* OpenCL Nvidia apps do exhibit this behavior, (and the WCG GPU app certainly does) under the right conditions you can create an Nvidia OpenCL app that does not use excessive CPU time. Proving this is simple: I'll show you such an app. Go run PrimeGrid's GPU version of the WW (Wall-Sun-Sun / Wieferiich) app on an Nvidia GPU in a Windows 10 machine. It will use almost no CPU. It is true that you have to write the app specifically to use the method that doesn't hog a CPU core, and you have to have the correct drivers, and it only works on some operating systems. It won't, for example, work on Windows 7. But Windows 10 does support it. Any developer who is interested in the nitty gritty details should contact me on PrimeGrid's Discord server and I'll hook you up with that app's developer. |
||
|
|
Ian-n-Steve C.
Senior Cruncher United States Joined: May 15, 2020 Post Count: 180 Status: Offline Project Badges: 
|
that's nice to hear that at least one project has employed such a method. too bad it's on a number theory project.
----------------------------------------do you know if there was any performance loss in coding it this way? I know back with SETI, the linux CUDA application that some of us ran, could be run in two configurations, normal, which had low CPU utilization, or with a command line parameter to force the full core usage. the latter ran faster. it's been my experience that projects tend to be adverse to app improvements for speed only. they care more about getting correct results than getting fast results. they are usually short staffed and without sufficient resources to make improvements solely in optimization. usually only creating new apps for new data types, or fixing bugs or known problems. The aforementioned SETI CUDA application was a grassroots effort by a talented CUDA developer who made the application in his spare time. and it was 4-5x faster than the project-provided app.  EPYC 7V12 / [5] RTX A4000 EPYC 7B12 / [5] RTX 3080Ti + [2] RTX 2080Ti EPYC 7B12 / [6] RTX 3070Ti + [2] RTX 3060 [2] EPYC 7642 / [2] RTX 2080Ti |
||
|
|
Michael Goetz
Cruncher United States Joined: Dec 11, 2017 Post Count: 35 Status: Offline Project Badges:
|
that's nice to hear that at least one project has employed such a method. too bad it's on a number theory project. do you know if there was any performance loss in coding it this way? I'm sure there has to be some loss of performance, and it might be able to measured if you tried hard enough, but there's certainly no noticeable difference in performance to the naked eye. |
||
|
|
Michael Goetz
Cruncher United States Joined: Dec 11, 2017 Post Count: 35 Status: Offline Project Badges:
|
it's been my experience that projects tend to be adverse to app improvements for speed only. they care more about getting correct results than getting fast results. they are usually short staffed and without sufficient resources to make improvements solely in optimization. usually only creating new apps for new data types, or fixing bugs or known problems. The aforementioned SETI CUDA application was a grassroots effort by a talented CUDA developer who made the application in his spare time. and it was 4-5x faster than the project-provided app. "Not Invented Here Syndrome" That's one advantage I have that SETI doesn't: In that situation, I'll offer the "outside" developer a "job" for the same compensation as the boss. I.e., nothing. :P If someone has an idea for a better mousetrap, we won't turn them away because they're "not one of us". We recently had a game-changing breakthrough in our software design / number theory that allowed us to essentially eliminate double checking of tasks while still being 100% certain that the task results were correct and didn't suffer from computation errors, software bugs, or cheating. Essentially, this doubled the project's throughput on many of our sub-projects. This idea (and the new software) came from a talented user. And, yes, we care more about accuracy than about speed. But we like speed too. :) |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 146 Status: Offline Project Badges: 
|
Validation is almost caught up. Sitting at 6k work units behind now. So it is catching up nicely. Also, I have kicked off the job to build more batches. We are crunching through these batches nicely. Thanks again for all the help on the stress test. It is truly amazing how much power there is available. Thanks, -Uplinger You are welcome! I was excited when I heard it was happening and set right to making sure all my computers were ready.  Does that mean I should start getting nvidia work units again? Earlier in the stress test I was running 4-5 at a time, now I have only gotten 2 in the last 2 hours.  |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 146 Status: Offline Project Badges:
|
Validation is almost caught up. Sitting at 6k work units behind now. So it is catching up nicely. Also, I have kicked off the job to build more batches. We are crunching through these batches nicely. Thanks again for all the help on the stress test. It is truly amazing how much power there is available. Thanks, -Uplinger Where would you like the pizza delivered?  Seriously though, is there a pizza fund I could contribute to? I would be happy to buy pizza. Or maybe after this stress test Uplinger will have a fund for new servers instead.  |
||
|
|
zdnko
Senior Cruncher Joined: Dec 1, 2005 Post Count: 229 Status: Offline Project Badges:
|
Does that mean I should start getting nvidia work units again? Earlier in the stress test I was running 4-5 at a time, now I have only gotten 2 in the last 2 hours. I have 49 intel wu and only one nvdia  EDIT: Maybe the scheduler reads the forum: I just got 18 nvdia wu.  [Edit 1 times, last edit by zdnko at Apr 27, 2021 9:51:25 PM] |
||
|
|
MindCrimeZ
Cruncher Joined: Feb 28, 2014 Post Count: 9 Status: Offline Project Badges:
|
Does that mean I should start getting nvidia work units again? Earlier in the stress test I was running 4-5 at a time, now I have only gotten 2 in the last 2 hours. I have 49 intel wu and only one nvdia EDIT: Maybe the scheduler reads the forum: I just got 18 nvdia wu. it does, yesterday I said i still haven't seen 1 intel gpu WU, today i got ...one. |
||
|
|
Former Member
Cruncher Joined: May 22, 2018 Post Count: 0 Status: Offline |
Right now i have 0 GPU work units and have not received any in the last 20 hours so are the GPU work units still going out to the people
|
||
|
|
kittyman
Advanced Cruncher Joined: May 14, 2020 Post Count: 140 Status: Offline Project Badges: 
|
The kitties are getting replacements for the all of the WUs they crunch, so I am not aware of any problems presently here.
----------------------------------------Meow  |
||
|
|
|