Advanced search

Forums : Technical Support : slow gpu
Message board moderation

To post messages, you must log in.

AuthorMessage
Pasi Nevalainen

Send message
Joined: 21 Feb 15
Posts: 6
Credit: 756,597
RAC: 54
Message 20636 - Posted: 13 Dec 2015, 17:54:21 UTC

My gpu(nvidia gtx 680) run too slow(only~50% normal speed) when i running to camb_boincc2docker. What is wrong ? If i run lhc,einstein or camb_legasy , nvidia works normally.
ID: 20636 · Report as offensive     Reply Quote
Profile Marius
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 29 Jun 15
Posts: 470
Credit: 4,276
RAC: 0
Message 20638 - Posted: 15 Dec 2015, 0:22:34 UTC - in response to Message 20636.  
Last modified: 15 Dec 2015, 0:23:34 UTC

I know very little about GPU computing, but my basic understanding is that along with the GPU you have one processor controlling the computation as well. Because C@H jobs by default will automatically use all processors available on the machine, its possible they're slowing down that controlling processor which in turn slows down the GPU. I see two solutions,

  • limit the number of CPU's that C@H jobs use
  • don't run C@H jobs at the same time as other jobs, instead run one then the other (this may be harder, I'm not sure the BOINC client can even do this automatically)

ID: 20638 · Report as offensive     Reply Quote
Profile Steve Hawker*

Send message
Joined: 10 Feb 13
Posts: 9
Credit: 244,929
RAC: 0
Message 20639 - Posted: 15 Dec 2015, 0:58:10 UTC - in response to Message 20636.  

My gpu(nvidia gtx 680) run too slow(only~50% normal speed) when i running to camb_boincc2docker. What is wrong ? If i run lhc,einstein or camb_legasy , nvidia works normally.


It sounds to me like you need to assign a full CPU core to your GPU. You can do this by a number of methods. The easiest might be to reduce the number of processors used in BOINC Manager preferences.

computing preferences >> processor usage

In the box for "On multiprocessor systems, use at most" put the percentage that uses all but one core. So if you have an eight core processor put 87.5%, a four core 75%, etc etc.

If that doesnt work, you can force camb_boincc2docker to use all but one processor. See http://www.cosmologyathome.org/forum_thread.php?id=7346&postid=20620#20620. Change avg_ncpus to meet your needs. I use it to make C@H use just one core but I'm running lots of other stuff too.

If you still have a speed problem then it isnt anything to do with C@H.
ID: 20639 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 17 Nov 14
Posts: 134
Credit: 5,412,499
RAC: 530
Message 20640 - Posted: 15 Dec 2015, 2:54:27 UTC

In addition to restricting the number of CPU cores that BOINC can use, I also found that I needed to reduce the priority of camb_boinc2docker using "Prio". reduce CPU priority

(That should probably be posted as a topic in this section for setting up you PC, rather than the "News" section.)
ID: 20640 · Report as offensive     Reply Quote
Profile Marius
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 29 Jun 15
Posts: 470
Credit: 4,276
RAC: 0
Message 20644 - Posted: 17 Dec 2015, 11:34:30 UTC - in response to Message 20640.  

In addition to restricting the number of CPU cores that BOINC can use, I also found that I needed to reduce the priority of camb_boinc2docker using "Prio". reduce CPU priority

(That should probably be posted as a topic in this section for setting up you PC, rather than the "News" section.)

So restricting the number of CPU cores wasn't enough? You were still seeing lagging, etc... after that?

Also note, the reduce CPU priority fix won't work for everyone as it has to do with how BOINC was installed. See here.
ID: 20644 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 17 Nov 14
Posts: 134
Credit: 5,412,499
RAC: 530
Message 20645 - Posted: 17 Dec 2015, 15:18:40 UTC - in response to Message 20644.  
Last modified: 17 Dec 2015, 15:20:39 UTC

So restricting the number of CPU cores wasn't enough? You were still seeing lagging, etc... after that?

I don't use my GPU cards (two GTX 750 Tis) for desktop purposes, only for Einstein BRP6 work. What I saw was an increase in the time required to finish a work unit. By reducing the CPU priority to "Low", I got the time back to (almost) what it was before. But I manage that PC remotely over the LAN, and don't even have a monitor connected to it. So I can't really make a good comparison for lag, since they all lag to some extent due to TightVNC over the LAN.

(That link to Berkley requires a password, but I just did a normal non-service install of BOINC, without the screen saver and with permissions to all users).
ID: 20645 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 17 Nov 14
Posts: 134
Credit: 5,412,499
RAC: 530
Message 20646 - Posted: 17 Dec 2015, 19:17:46 UTC - in response to Message 20645.  

But maybe I did not answer your question directly. First I reserved two cores for the GPU by using the special "app_config.xml" noted previously, but that was not enough. Maybe it would work to eliminate desktop lag, but I could still see the differences on the Einstein work units. I had to reduce the CPU priority using Prio to fix that (in addition to reserving the two cores).
ID: 20646 · Report as offensive     Reply Quote
Profile Marius
Project administrator
Project developer
Project scientist
Avatar

Send message
Joined: 29 Jun 15
Posts: 470
Credit: 4,276
RAC: 0
Message 20647 - Posted: 18 Dec 2015, 11:43:08 UTC - in response to Message 20646.  
Last modified: 18 Dec 2015, 11:43:40 UTC

Ah OK, thanks for the clarification. That's really bizarre though. Despite two free cores and the fact that the C@H job running on the others doesn't touch your GPU, it was still slowing down a GPU job? Any theories why that was?


Btw the link was this:
Rom Walton: I have not fiddled with VboxSvc's process priority. On Windows, we only launch it if BOINC was installed as a service. Otherwise, Windows launches it with the DCOM Service and its process priority is out of our hands.

So yes, it basically depends on whether you installed BOINC as a service on Windows or not. Would you be willing to write up a short description of the steps you had to take to reduce the CPU priority, so that I might include it in the FAQ?
ID: 20647 · Report as offensive     Reply Quote
Larry

Send message
Joined: 20 Oct 13
Posts: 10
Credit: 195,064
RAC: 0
Message 20657 - Posted: 26 Dec 2015, 18:45:42 UTC

So..., what is the procedure for using "Prio" to set the process of VBoxSVC to "Low?"
ID: 20657 · Report as offensive     Reply Quote
Jim1348

Send message
Joined: 17 Nov 14
Posts: 134
Credit: 5,412,499
RAC: 530
Message 20658 - Posted: 26 Dec 2015, 20:03:51 UTC - in response to Message 20657.  

So..., what is the procedure for using "Prio" to set the process of VBoxSVC to "Low?"

First install Prio and reboot. Then open up Windows Task Manager (right-click on the taskbar), and in the "Processes" tab you will see VBoxSVC; right-click on that and set the priority to "Low" and reboot. That should fix all the VBoxHeadless tasks to "Low" also.
ID: 20658 · Report as offensive     Reply Quote
Crystal Pellet

Send message
Joined: 12 Feb 13
Posts: 23
Credit: 363,354
RAC: 0
Message 20668 - Posted: 29 Dec 2015, 21:30:39 UTC - in response to Message 20646.  

I had to reduce the CPU priority using Prio to fix that (in addition to reserving the two cores).

I got unexpected troubles using Prio - Process Priority Control (v2.0.0.2960) for lowering the cpu-priority.

When running more VM's of several BOINC projects (I had 1 CMS-dev, 2 ATLAS and 1 Cosmo) most VM's are stopping and the task in BOINC is postponed for 24 hours to retry after that period.

2015-12-29 19:27:25 (676): VM state change detected. (old = 'running', new = 'gurumeditation')
2015-12-29 19:27:35 (676): NOTE: VirtualBox has failed to allocate enough memory to continue.
2015-12-29 19:27:35 (676): This might be a temporary problem and so this job will be rescheduled for another time.
2015-12-29 19:27:35 (676): Powering off VM.


Free memory at that moment was 13GB.
This error was reproducible and disappeared after uninstalling 'PRIO' from my win7 machine.
Lowering the CPU-priority by hand or with eFMer's program priority, I didn't had that issue (could run up to 8 VM's).
Disadvantage of last mentioned program is the fact, that all programs only can get the same priority.
ID: 20668 · Report as offensive     Reply Quote

Forums : Technical Support : slow gpu