Message boards :
Number crunching :
MilkyWay CUDA takes up a whole core
Message board moderation
Author | Message |
---|---|
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
I noticed that MilkyWay CUDA takes up almost a whole core as you can see in this image (the 'img' tag didn't work with windows-skydrive). Anybody else that have seen this behaviour? When another MilkyWay task runs on this system (it only runs the GPU tasks) I'll check it again too see if it's a recurring problem. Anthony. The longer I live, the more reasons I develop for wanting to die. |
Send message Joined: 8 May 10 Posts: 576 Credit: 15,979,383 RAC: 0 |
It's a known recent Nvidia driver problem where OpenCL things take up huge amounts of CPU time for no reason. |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
|
Send message Joined: 10 Oct 11 Posts: 1 Credit: 278,285 RAC: 0 |
I'm having the exact same issue. I disallow all my projects to use CPU time, as they have full use of my GPU. Yet as you mention above it has lead us to nvidia driver bug? I'm on the latest 285.62 - Anyone seen or heard of any updates to this? |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
|
Send message Joined: 17 Mar 09 Posts: 7 Credit: 1,563,090 RAC: 0 |
280.13 shows this effect on all WU, taking 200% on 2 cores of my 2500k |
Send message Joined: 4 May 08 Posts: 1 Credit: 4,919,985 RAC: 0 |
I had the same problem with nvidia driver 280 and 285. CUDA takes at least one whole core (if the cpu idles CUDA even takes 2 cores). There are no problems with driver 266. |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
|
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
I just wrote a support request on NVIDIA's website about this problem and another problem I have (driver crashes when suspending Collatz on my GF-8600GT, that's also in my system together with my GTX-550Ti; my GTX-550Ti runs al other GPU projects I have). Anthony. The longer I live, the more reasons I develop for wanting to die. |
Send message Joined: 30 Jan 09 Posts: 21 Credit: 13,256,888 RAC: 0 |
I had the same problem with nvidia driver 280 and 285. CUDA takes at least one whole core (if the cpu idles CUDA even takes 2 cores). +1 Is there anything that could be changed in MW app to fix this? |
Send message Joined: 14 Feb 09 Posts: 999 Credit: 74,932,619 RAC: 0 |
I had the same problem with nvidia driver 280 and 285. CUDA takes at least one whole core (if the cpu idles CUDA even takes 2 cores). Not as long as it is a OpenCL app. |
Send message Joined: 8 May 10 Posts: 576 Credit: 15,979,383 RAC: 0 |
I'm pretty sure I can write a workaround for it.I had the same problem with nvidia driver 280 and 285. CUDA takes at least one whole core (if the cpu idles CUDA even takes 2 cores). |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
That would be great :) BTW, how much memory does a GPU WU take, because maybe I could run more than 1 WU on a 1GB GTX-550Ti. But then I'm gonna need an app_info.xml also for it. On SETI and Einstein I already run 3 WU's concurrently. Thanks in advance for the rewriting. Anthony. The longer I live, the more reasons I develop for wanting to die. |
Send message Joined: 8 May 10 Posts: 576 Credit: 15,979,383 RAC: 0 |
BTW, how much memory does a GPU WU take, because maybe I could run more than 1 WU on a 1GB GTX-550Ti. It should take (mu * r * (number of streams + 1)) * (1 or 2) * sizeof(double) bytes. For a typical work unit this would be (1600 * 1400 * (3 + 1)) * 8 = 68MB. |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
|
Send message Joined: 14 Feb 09 Posts: 999 Credit: 74,932,619 RAC: 0 |
You also have to take into account how much of a load 1 WU has on the GPU, if you are already around 80 to 90 percent, 2 at a time would be the max. Otherwise you are slowing down all wu needlessly. This will have the appropriate app_info to start. http://www.arkayn.us/forum/index.php?action=downloads;sa=view;down=33 |
Send message Joined: 15 Jul 08 Posts: 383 Credit: 729,293,740 RAC: 0 |
It's possible to achieve a small benefit by running 2 WUs per GPU because there is a small pause in GPU usage each time a WU nears completion. If 2 WUs are run in tandem, then the second one will use all available GPU during this period. Running more than 2 at a time will yeid no additional benefit and has the potential to cause additional problems. |
Send message Joined: 9 Dec 07 Posts: 65 Credit: 8,015,709 RAC: 0 |
|
Send message Joined: 2 Feb 08 Posts: 3 Credit: 6,013,119 RAC: 0 |
the GPU-apps could be much, much longer then only a few minutes. it´s not good to share so short WU. that´s a heavy, heavy load for the server(s) even a lot of WU are returning in minute takting :/ please increase the calculation-time for 1a. ATI 1b. Nvidia up to 50x-100x and the servers and databases will be very stable and can do the work step-by-step without trouble ;) Norman |
Send message Joined: 30 Jan 09 Posts: 21 Credit: 13,256,888 RAC: 0 |
I'm pretty sure I can write a workaround for it. And I'm pretty sure you can do it. But... are you doing it? |
©2024 Astroinformatics Group