Using GPU - questions
10 Jul 2016 4:09:27 UTC
Topic 198681
(moderation:
Haven't been able to find the answer(s).
When Boinc uses the GPU:
1. does it use a percentage of the cores? If my card has 200 cores does it use 1 or (say) 100?
2. how many jobs can the GPU handle at one time?
3. configuration - anything I can do to cause more GPU usage or is it fixed?
Thanks
Using GPU - questions
)
1)It will use as much of the gpu as the workunit is designed too use, but it will try to use all of the gpu cores it can. In fact that's one thing that helps makes one gpu faster than another one, more gpu cores on it.
2)Your gpu may be able to handle 2 or 3 workunits at a time but only experimenting will tell you for sure.
3)Yes you will need a config file to do that, it's not hard at all.
FIRST get your gpu crunching 1 workunit at a time, I see you have some you are working on now, and then get a program that can tell you how much of your gpu that workunit is using. In Windows I use GPU-Z or MSI's Afterburner, I don't know what works on Linux machines though. Those programs can tell you how much load the workunit is putting on your gpu, which you can then use to figure out if running another workunit at the same time on it would be good or bad. Some gpu apps are designed to max out the gpu and just finish much quicker so adding more workunits to run at the same time actually slows things down.
What you are looking for is something in the 90%+ percent usage of your gpu with your workunits to get the max out of it. That being said the more workunits you run the more heat your gpu will produce and too much heat can destroy a video card, running Boinc does not give our gpu's a break like gaming can. Gamers may do a 20 hour session but then they stop and the gpu gets a chance to cool down, with Boinc there are no breaks it's full speed 24/7/365!! What I'm getting at is GOOD airflow is ESSENTIAL to gpu crunching!!
Try this webpage for GPU-Z or Afterburner or perhaps find some Linux alternatives: http://www.guru3d.com/files-categories/videocards-handy-utilities.html
1st, thanks for the
)
1st, thanks for the reply.
OK, (running FC22) google... you want nvidia-smi, search, search, ah download nvidia.xyz, run, error out, search, search, "... in included in xorg..nvid..." WT?, locate nvidia-smi, -> /usr/bin/nvidia-smi - there goes 1/2 hr I'll never get back :-/ was already there.
So, ran: "nvidia-smi -l 1" and 2 minutes later (bmgr says don't use gpu for 2 min of non-keyboard usage) and got the following:
-------------------------------------------------
| GPU PID Process name Usage | |=============================================================================| | 0 Not Supported | +-----------------------------------------------------------------------------+ Sun Jul 10 07:02:50 2016 +------------------------------------------------------+ | NVIDIA-SMI 340.96 Driver Version: 340.96 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | |===============================+======================+======================| | 0 GeForce GTS 250 Off | 0000:05:00.0 N/A | N/A | | 83% 83C P0 N/A / N/A | 291MiB / 1023MiB | N/A Default | +-------------------------------+----------------------+----------------------+ +-----------------------------------------------------------------------------+ | Compute processes: GPU Memory | | GPU PID Process name Usage | |=============================================================================| | 0 Not Supported | +-----------------------------------------------------------------------------+___________________________________________
The only thing I got out of the above was the fan was 40% & 50C for the first 2 min then began to climb.
Here's what started this. I was running FC19 and averaging right at 30,000 Boinc credits per day. It ain't broke so let's upgrade to FC22. Average dropped to about 5,000/day. Via Google: Oh the GTS 250 is old and not supported any more (but there's a legacy driver) - some time later, much forehead blood on screen, boincmgr says "seti..... cuda60". Suc-Sess! and my per day has climbed to 7,000/day ... (more like SUCKsess).
Aside: I've been on the page for an hr+ now - playing ([---] Resuming GPU computation) I can't get systemctl start boinc-client.service to work. It starts, runs for a few seconds and poof - gone. I'll keep slog'n. Any other words of advice?
Thanks
RE: 1st, thanks for the
)
Turn off the wait 2 minutes before starting to crunch setting until you get it fixed, that way things will start RIGHT NOW instead of making a change, waiting, making another change, waiting, etc, etc.
As for the actual "get systemctl..." stuff I have no clue, I'm a Windows guy and that command line stuff isn't needed.
And yes the 250 IS an older gpu, but as you said something is better than nothing. With the new 1080, 1070 and 1060 models all coming out now you should be able to find some very good prices on some older, but still an upgrade for you, models shortly. I'm running GTX 760's here and they work just fine, I'm only running one unit at a time though with the default settings. I know they could do better but it's 90+F in my basement and more units means more heat and I don't need anymore of that right now.
Thanks I'll give it a try.
)
Thanks I'll give it a try. Yes my 250 is an older gpu - but... before I upgraded to FC22 I was getting ~30,000 credits/day - now 7,000. Same board, different SW - where would you look to solve the problem? (rhetoric).