Printable View
Previous to upgrading to Boinc 7.0.25 CPU usage for ATI wu's on MW was around .05. I'm now seeing .9+ cpu per gpu. Since I have 4 ATI cards in this rig and a quad core, MW now uses all four cores to work with the gpu's. Anyone else seeing this. Is it something that changed in the upgraded Boinc Manager or did MW change their WU. I use to like to run mw beside my CPU crunching, but can't now.
No, that is dictated by the project. You can override it (at least on MilkWay) using an app_info.xml file but I do not recall the specifics. I look for one of my old ones and see if I can edit it for the current app.Quote:
Originally Posted by Al McAdams
A lot of that depends upon the app and the algorithms used in the app. OpenCL is very prone to using up to full CPU than CUDA or CAL/Brook+. As the number of items increases per kernel, the amount of CPU increases. CAL/Brook+ can have 10 times as many items per kernel (a kernel is a subroutine that is run on the GPU) and not use any CPU. More items = higher GPU utilization in OpenCL. Finding the balance where it keeps the GPU busy without using CPU time is difficult.
For some algorithms, they spend more time loading the memory and reading it back again than they do actually crunching and that uses more CPU as well. So, algorithms that don't run in parallel very well tend to use a lot more CPU in order to keep the GPU busy. That's what I consider sticking a square peg in a round hole. Just because you can make an app run on a GPU doesn't necessarily mean it should run on a GPU. Chances are, those GPU apps probably only run 8-10 times faster than the CPU apps whereas those which can be easily run in parallel can achieve speeds of 100 times faster on GPUs.
I've found that in addition to MW taking 4 cores of my quad, so does Donate and DiRT ATI takes 3 of 4. This is sort of a noob question, so bear with me. I don't like using App_info files because I usually screw those up, but I have found that using <ncpus></ncpus> lines in the cc_config file works too. DiRT is my example. Running 4 ATI work units takes 3 cores of cpu. Load on the cores varies all over the place, from 6% to 54%. By adding <ncpus>7</ncpus> to the config file I get 100% usage on all cpu cores and 99% usage on 3 gpus and one gpu @94%. This seems too easy a fix and I'm sure it negatively effects wu times to some degree, so tell me what I'm missing in my logic.
Attachment 819
I think the issue you will run into is the gpu wont be getting max credit in this format but if you are wanting more out of the cpu specially during challenges that is one option to use.. Personally for me i swap over to moo on the AMD cause of the low cpu usage and still get decent gpu credits without loosing to much CPU or doing anything to the config file but i still havent made the move to 7.0xx, not sure which project will lead to less cpu usage on the nvidia side.
You were right on that one. I tried a few Donate wus and with the ncpus set to 7 (100% cpu load) the gpus were starved and the Donate's took 2300 sec to complete. I dropped the ncpus back to 6, essentially freeing up a core and the Donate's came in at 1100 to 1200 sec. This is about the same time it took when Donate was taking all 4 cores, so I think I've found a sweet spot for my system. Later today I'll try some MWay wus because I have a much better sense of what my gpus were capable of prior to the increased cpu usage. I'll let you know how that goes.Quote:
Originally Posted by coronicus
Edit: Seeing the same with MWay. GPU times about the same as prior, but know I have 3 cores to do cpu projects. I'm happy now!