Message boards :
Number crunching :
Concurrent GPU tasks
Message board moderation
Author | Message |
---|---|
Jason A. Countryman Send message Joined: 29 Aug 03 Posts: 139 Credit: 50,172,873 RAC: 2 |
Hello all, I am in the process of setting up a new app_config file for v8 and plan on running stock apps for a bit to get a benchmark. I know there was a post linking a tool to determine the best number of concurrent tasks for each GPU, but I can't seem to find it. I am running linux with an FX-8300 processor and a Radeon 5850 GPU. Right now I am running 3 GPU tasks at once, with each task getting .75 of a core. Any suggestions? Thank you. |
TBar Send message Joined: 22 May 99 Posts: 5204 Credit: 840,779,836 RAC: 2,768 |
I think the tool you're thinking about was Windows only, and hasn't been updated in a while. The best method is by observing the GPU load with different tasks. There is a handy tool for Linux here, http://lunatics.kwsn.info/index.php?action=downloads;sa=view;down=460 With my Cayman I found by running 2 MB tasks the GPU load would go from around 90% to the high 90s and I would save about 20 seconds over 6 minutes and around a minute over 16 minutes. Unfortunately running 2 MBs at a time introduced noticeable Screen Lag whereas running One task didn't. I choose to run One task with two free cores, seems the times were a few seconds slower with just One free core. You might want to check the GPU loads with the tool and see when your GPU reaches the high 90s, increasing the number of tasks once the GPU is in the high 90s is usually not helpful. I see you have one of those Mac nVidia GPUs that doesn't work very well with the OpenCL App. There is a New Mac CUDA App that would probably work well with that GPU. You would have to switch to Anonymous platform to try it. In your case you would just install the latest CUDA driver, place the CUDA42 files in your setiathome.berkeley.edu folder, and then reinstall BOINC to set the file permissions. To switch back you would just remove the files or simply rename the app_info.xml file so BOINC wouldn't use it. You would also want to run down your existing tasks first, to keep from creating Ghosts. The App is here, http://www.arkayn.us/forum/index.php?topic=191.msg4411#msg4411 |
Jason A. Countryman Send message Joined: 29 Aug 03 Posts: 139 Credit: 50,172,873 RAC: 2 |
Thanks for the info, I will try the CUDA app for the Mac. I have been checking on the GPU utilization using the aticonfig commands on the linux box. It was around 90-94% utilization with one task, and 97-99% when running 3 tasks. If it stays that high with only 2 tasks, should I leave it at 2? I don't mind the screen lag as this is primarily a web/media server and I don't use it directly very often. |
TBar Send message Joined: 22 May 99 Posts: 5204 Credit: 840,779,836 RAC: 2,768 |
If you were seeing around 90-94% load with just One task I wouldn't run any more than 2. With 2 you would see very little improvement and with 3 it would probably be slower. The Mac app comes with a CPU App that will probably be about the same as your current results. Just in case, you might want to copy your stock Apps to another folder before you make the change just in case you want to go back to the stock CPU App. On my machine the sse41 App is better than the stock CPU App. |
Jason A. Countryman Send message Joined: 29 Aug 03 Posts: 139 Credit: 50,172,873 RAC: 2 |
So I changed it to .5 GPU and 1.5 CPU in app_config. I am now running at 96-97% GPU utilization, with 5 of 8 cores also running tasks. Im not sure if I should give it more CPU cores or go back to 3 tasks. Update: I tried going back to 3 GPU tasks, and giving each task 2 cores. The GPU utilization would drop into 80% sometimes, then go back to 98-99%. I went back to 2 tasks with 1.5 cores for each. That seems relatively stable at 96-97%. |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13720 Credit: 208,696,464 RAC: 304 |
So I changed it to .5 GPU and 1.5 CPU in app_config. My Windows system is able to run 2 GTX 750Tis with 2 WU on each with no problems running work on all CPU cores (physical and virtual). .04 CPU is generally plenty for CUDA work. Grant Darwin NT |
rob smith Send message Joined: 7 Mar 03 Posts: 22158 Credit: 416,307,556 RAC: 380 |
Your GPU term is correct, but the CPU term does not need to b any higher than about 0.75 Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Rasputin42 Send message Joined: 25 Jul 08 Posts: 412 Credit: 5,834,661 RAC: 0 |
I understand, that the allocation of 0.5, 0.25 or so cpus is just an administrative figure for boinc to decide which and how many tasks to run. I has no meaning in terms of how much cpu percentage is actually used by a task. |
Jason A. Countryman Send message Joined: 29 Aug 03 Posts: 139 Credit: 50,172,873 RAC: 2 |
So I tried setting the CPUs to .75, and my utilization went down to the 80% range. I then set it to 1, and I'm back to 94-95%. Im wondering if its worth loosing another core to get to 97% or best to leave it where it is. |
rob smith Send message Joined: 7 Mar 03 Posts: 22158 Credit: 416,307,556 RAC: 380 |
what utilisation are you talking about - CPU or GPU? Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Jason A. Countryman Send message Joined: 29 Aug 03 Posts: 139 Credit: 50,172,873 RAC: 2 |
GPU utilization |
©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.