| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: GPU Support Forum
Thread: HCC GPU - Thread Priority should be increased (to 6) to not starve the GPU |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 28
|
|
| Author |
|
|
nanoprobe
Master Cruncher Classified Joined: Aug 29, 2008 Post Count: 2998 Status: Offline Project Badges: 
|
You'll never get full GPU load on Nvidia cards that are running a project that runs on the OpenCl platform. I can't make it any simpler than that. There is no way to do it. GPUGrid runs on the proprietary CUDA platform. That is why ATI cards don't work there. You can't compare results from there to here. It's like apples and oranges.
----------------------------------------
In 1969 I took an oath to defend and protect the U S Constitution against all enemies, both foreign and Domestic. There was no expiration date.
----------------------------------------  [Edit 1 times, last edit by nanoprobe at Apr 9, 2013 12:40:18 AM] |
||
|
|
Jacob Klein
Cruncher Joined: May 31, 2007 Post Count: 28 Status: Offline Project Badges: 
|
Fair enough.
----------------------------------------But I'm hopeful to get the 97-98% constant load that I see when the CPU is not overloaded with CPU tasks... to be applied to the case where the CPU IS overloaded with CPU tasks. And I believe a priority change is key to achieving it. That's what I'm asking for here. [Edit 2 times, last edit by Jacob Klein at Apr 8, 2013 4:03:56 AM] |
||
|
|
nanoprobe
Master Cruncher Classified Joined: Aug 29, 2008 Post Count: 2998 Status: Offline Project Badges:
|
*scratches head*
----------------------------------------What part of you'll never get a constant 97-98% GPU load on an OpenCl project with an Nvidia card is it that you don't understand? It cannot be done until Nvidia does something about their OpenCl driver support.
In 1969 I took an oath to defend and protect the U S Constitution against all enemies, both foreign and Domestic. There was no expiration date.
|
||
|
|
Jacob Klein
Cruncher Joined: May 31, 2007 Post Count: 28 Status: Offline Project Badges:
|
I see you edited your earlier post, so it doesn't say "Last Post" :) Nice.
----------------------------------------Look, bud, I'm not sure if you're actually reading what I'm writing, or not. I assume you are. But I'll repeat it. Here goes. Please read it, and please understand. On my 8-CPU machine, on my GTX 460: - when the only thing running on the entire system is a single WCG HCC GPU task, it runs at about 94-96% GPU Load. - when I use app_config with <gpu_usage> of 0.5 (to run 2 of those tasks on that GPU), then with only those 2 GPU tasks running, the GPU absolutely runs at 98% load, constant. If the number of CPU-intensive processes exceeds the number of CPUs on the system, then currently, the wcg_hcc1* processes can get swapped out of the CPU, because the process currently runs at the same priority as the other CPU processes. If the number of CPU-intensive processes is equal to or lower than the number of CPUs on the system, then none of the CPU-intensive processes get swapped out, and I can get that 98% load. Some GPU tasks take fractional CPU, and BOINC correctly slightly overloads the system to keep the system full loaded, during the times when those GPU tasks may not be using much of the CPU. But... if the CPU process of a GPU task gets swapped out when the GPU task actually DOES require the CPU, the result is that GPU performance suffers. GPUGrid solved the problem by increasing the process priority to 6, and increasing the cpu-intensive thread within that process, to 6 or 7. In our example, if my system is fully loaded (2 WCG tasks on the GTX460 each taking 1 CPU, 2 GPUGrid tasks on the GTX660Ti each taking fractional CPU, 6 CPU Tasks)... The GPU Load on the GTX460 currently varies greatly, as the processes get cycled out of the CPU. The WCG HCC GPU tasks' performance unnecessarily suffers. If we increased the process priority and cpu-intensive-thread priority, then for my fully-loaded scenario... I am positive we could maintain the 98% GPU Load that I currently get when the system is not fully loaded. Do you get it yet? [Edit 5 times, last edit by Jacob Klein at Apr 9, 2013 4:26:12 AM] |
||
|
|
Jacob Klein
Cruncher Joined: May 31, 2007 Post Count: 28 Status: Offline Project Badges:
|
knreed,
I understand you guys are wrapping up the HCC GPU run. Please, especially for future GPU projects, please consider implementing what I am suggesting, to keep GPU Load up and keep task throughput optimal. Thank you, Jacob Klein |
||
|
|
BladeD
Ace Cruncher USA Joined: Nov 17, 2004 Post Count: 28976 Status: Offline Project Badges: 
|
knreed, I understand you guys are wrapping up the HCC GPU run. Please, especially for future GPU projects, please consider implementing what I am suggesting, to keep GPU Load up and keep task throughput optimal. Thank you, Jacob Klein Since future GPU projects will be completely different, it may not be needed. Also, the Nvidia drivers should be better by then.  But is there an OpenCl project with a Nvidia card that get a constant 97-98% GPU load? |
||
|
|
Jacob Klein
Cruncher Joined: May 31, 2007 Post Count: 28 Status: Offline Project Badges:
|
Help Conquer Cancer, when configured for 2-tasks-on-1-NVIDIA GPU, does get me constant 98% GPU load. Is it OpenCL?
|
||
|
|
knreed
Former World Community Grid Tech Joined: Nov 8, 2004 Post Count: 4504 Status: Offline Project Badges:
|
We have ~3 weeks or so remaining in the project. As a result, we are not going to make changes to the apps at this point. We will examine these issues going forward and also discuss with BOINC to make sure the optimal behavior is set by the client primarily and only relying on project apps doing things if they must.
|
||
|
|
Ingleside
Veteran Cruncher Norway Joined: Nov 19, 2005 Post Count: 974 Status: Offline Project Badges: 
|
You'll never get full GPU load on Nvidia cards that are running a project that runs on the OpenCl platform. I can't make it any simpler than that. There is no way to do it. Well, going by various posts over at the Folding@Home-forums about the new Core-17 currently in beta-testing, this new core is OpenCL-only, and various posts from Nvidia-users claims 98-99% GPU-usage (example GTX-460 and GTX 675M). Where's also a mentioning some CPU-code to ocassionally verify the progress was added to the beta, so during this verifying the GPU-usage drops. No idea how accurate the 98-99% GPU-usage-claims are, and neither how large and frequent the drops due to the debug-code are. Since I've not got any Nvidia-cards someone else would need to test this for themselves... Oh, and while didn't explisitely state it, in Folding@Home users is normally only running one wu per GPU.  "I make so many mistakes. But then just think of all the mistakes I don't make, although I might." [Edit 1 times, last edit by Ingleside at Apr 9, 2013 9:50:22 PM] |
||
|
|
Jacob Klein
Cruncher Joined: May 31, 2007 Post Count: 28 Status: Offline Project Badges:
|
Thank you knreed.
----------------------------------------I wouldn't have brought the issue up if I didn't think it would really help. And, as you've seen, I've tried to provide as much detail as possible in my experiences. GPUGrid.net used the priority to ensure that their GPU apps aren't starved by lack of CPU resources, so I hope you guys can do the same. I'm already looking forward to World Community Grid's next GPU project, whenever that may be! Thank you, Jacob Klein [Edit 2 times, last edit by Jacob Klein at Apr 9, 2013 11:07:50 PM] |
||
|
|
|