| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Support
Forum: GPU Support Forum
Thread: Intel HD GPU not useable though listed by CLinfo |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 38
|
|
| Author |
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 145 Status: Offline Project Badges: 
|
Yes it is wierd, in fact I would classify it as a bug. After we talked on the forum several days ago, I thought it made no sense. So, I set about trying to replicate your issue. I was successful and can reproduce it easily. We have been digging into it since then. It seems the scheduler only sends work for intel_gpu to machines that only have an intel_gpu if they have cpu computing enabled. I let my test machine build up about a days worth and then disabled cpu computing and it still had not gotten any more before it finished them. Is there any update on this? It seems that after initial burst of GPU WUs I'm now receiving only CPU WUs. I'm afraid that after stress test CPU WUs are only thing I'm gonna catch on those laptops The intel GPU tasks are available. Make sure that it is requesting tasks for CPU and Intel GPU. Have a max tasks set for OPN in the profile and cache set high enough that it will keep requesting CPU and Intel GPU after it hits the max tasks for OPN.  |
||
|
|
biini
Senior Cruncher Finland Joined: Jan 25, 2007 Post Count: 334 Status: Offline Project Badges: 
|
Thank you for your reply goben. I've set following on the Dell Latitude E7440 laptop profile
----------------------------------------- Available projects: opn - Allow Research to run on my CPU : YES - On multiprocessors at most use: 2 processors (2core, 4 thread cpu) - Graphics card usage: yes x4 - Cache 7.0 extra days of work Project limits - OPN 3 tasks It seems that I still only get CPU work units. Do the settings look ok ? E: On boinc client tasks tab I have 2 CPU tasks running and third one is ready to start ---------------------------------------- [Edit 2 times, last edit by biini at May 4, 2021 12:09:12 PM] |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 145 Status: Offline Project Badges:
|
Thank you for your reply goben. I've set following on the Dell Latitude E7440 laptop profile - Available projects: opn - Allow Research to run on my CPU : YES - On multiprocessors at most use: 2 processors (2core, 4 thread cpu) - Graphics card usage: yes x4 - Cache 7.0 extra days of work Project limits - OPN 3 tasks It seems that I still only get CPU work units. Does the settings look ok ? E: On boinc client tasks tab I have 2 CPU tasks running and third one is ready to start Yes, that looks fine, although 7 day cache is quite high. I normally have that set to 1-2 days at the most. Meaning X + Y is less than or equal to 1-2. So right now one profile is .5 + 0.05, and another is .75 + .1. Connect to network about every X days Cache Y extra days of work What is your log showing now? Maybe there is something useful there. Could you post a fetch cycle with sched_op_debug and work_fetch_debug on? At least from: [work_fetch] [World Community Grid] update requested by user to: [work_fetch] Request work fetch: RPC complete |
||
|
|
biini
Senior Cruncher Finland Joined: Jan 25, 2007 Post Count: 334 Status: Offline Project Badges:
|
04/05/2021 16:01:46 | World Community Grid | update requested by user ---------------------------------------- [Edit 1 times, last edit by biini at May 4, 2021 1:03:43 PM] |
||
|
|
biini
Senior Cruncher Finland Joined: Jan 25, 2007 Post Count: 334 Status: Offline Project Badges:
|
Ok, found a possible solution.
----------------------------------------It was (again?) Intel GPU Driver. Downgraded from 27.20.100.9466 to 27.20.100.8681 (CL3.0 -> CL2.1) and boinc started to download Intel GPU WUs immediatly. |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 145 Status: Offline Project Badges:
|
Ok, found a possible solution. It was (again?) Intel GPU Driver. Downgraded from 27.20.100.9466 to 27.20.100.8681 (CL3.0 -> CL2.1) and boinc started to download Intel GPU WUs immediatly. Great, I am glad you got it to work again.  |
||
|
|
highCon
Cruncher Joined: Jul 19, 2009 Post Count: 8 Status: Offline Project Badges:
|
I also have an Intel iGPU and can confirm that BOINC only downloads GPU-tasks when setting 'Allow research to run on my CPU?' to YES. As we know, this bug has the unfortunate downside of downloading & running CPU-tasks. However, I found a way to at least prevent the execution of those and only run GPU-tasks - it can be achieved over the almost-allmighty app_config.xml with the following content:
----------------------------------------
(Note: setting max_concurrent to 0 has no effect - apparently the default ('unlimited'); setting to 1 runs one CPU-task, etc.) Put it into BOINC\projects\www.worldcommunitygrid.org and in BOINC Manager trigger 'Options -> Read config files' and it'll immediately stop running CPU-tasks. In this way you don't need to manually suspend CPU-tasks and it may reduce the amount of babysitting the OPN project. I found it out an hour ago, so I still have to observe whether it automatically downloads new GPU-tasks while having many 'Ready to start' CPU-tasks in queue... [edit, the day later: yes, it works as intended: it automatically downloads new GPU-tasks once some are finished. In total, there are only as many GPU-tasks in queue as I have specified with the project limit for OpenPandemics in the device profile. At the same time there are as many CPU-tasks, but they are not processed/running. (As suggested by goben above, set the project limit to a low number to avoid having too much CPU-tasks in queue; BOINC Manager will make sure to fill up your queue with new GPU tasks anyway.) Also, don't forget to uncheck the option 'If there is no work available for the project(s) I have selected above, please send me work from another project.' - otherwise you'll get CPU tasks from other projects which will immediately run as the app_config.xml above only relates to CPU-tasks of OPN.] [Edit 1 times, last edit by highCon at May 6, 2021 7:34:01 AM] |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 145 Status: Offline Project Badges:
|
I also have an Intel iGPU and can confirm that BOINC only downloads GPU-tasks when setting 'Allow research to run on my CPU?' to YES. As we know, this bug has the unfortunate downside of downloading & running CPU-tasks. It is worse than that. It actually requires you to request CPU tasks, not just have it enabled. I was able to reproduce it with CPU enabled, but not requesting CPU tasks(from max_concurrent + use at most X% of CPUs). I found it out an hour ago, so I still have to observe whether it automatically downloads new GPU-tasks while having many 'Ready to start' CPU-tasks in queue... It can get tasks with OPN1 set to max_concurrent = -1. (Note: when setting max_concurrent to 1, it runs only 1 CPU-tasks; setting it to 0 has no effect - apparently the default.) Yes, 0 = no limit. This is why in the profile project limits section it goes: Unlimited, 1, 2, 3 etc. This is similar to some other settings such as "Suspend work if CPU usage is above" X "% of cpu" where "If you set this value to 0, it means no restriction is enforced." |
||
|
|
highCon
Cruncher Joined: Jul 19, 2009 Post Count: 8 Status: Offline Project Badges:
|
Hey goben, just to clarify: with downloading I mean requesting of course.
----------------------------------------I was able to reproduce it with CPU enabled, but not requesting CPU tasks(from max_concurrent + use at most X% of CPUs). Not sure what you mean. Can you clarify? It can get tasks with OPN1 set to max_concurrent = -1. Yes, you get them, but they don't run. Again, this is a dirty trick to circumvent the bug that you have found (in the sense that no CPU-tasks run if you don't want that). There is no way around to have them downloaded, though, simply because WCG appears to consider GPU-tasks related to 'research to run on CPU'. Yes, 0 = no limit. This is why in the profile project limits section it goes: Unlimited, 1, 2, 3 etc. This is similar to some other settings such as "Suspend work if CPU usage is above" X "% of cpu" where "If you set this value to 0, it means no restriction is enforced." Beware that the max_concurrent parameter in app_config.xml does not relate to the number of downloaded/in-queue tasks, but to the number or concurrently running tasks. Again, it's the only parameter that I found that can completely stop running any task. [Edit 1 times, last edit by highCon at May 6, 2021 8:04:41 AM] |
||
|
|
goben_2003
Advanced Cruncher Joined: Jun 16, 2006 Post Count: 145 Status: Offline Project Badges:
|
Hey goben, just to clarify: with downloading I mean requesting of course. I was able to reproduce it with CPU enabled, but not requesting CPU tasks(from max_concurrent + use at most X% of CPUs). Not sure what you mean. Can you clarify? Sure. It is the reason I made the distinction that you actually have to request cpu tasks, not just have it enabled. Meaning it has to say "Requesting new tasks for CPU and Intel GPU", not "Requesting new tasks for Intel GPU" to receive intel gpu tasks. When I set CPU enabled, but had max_concurrent for OPN1 and project_max_concurrent to 2(I was running 2 intel GPU units) along with "use at most 12% of CPUs"(The CPU has 4core/8threads and avg_ncpus was set to 0.5 for intel GPU units): It did not request CPU tasks due to MC(max concurrent) shortfall. Meaning work_fetch_debug showed using MC shortfall of 0 instead of shortfall (number > 0) for the CPU, so the client did not actually request CPU tasks. I hope that helped clear it up some, if not hopefully it will be more clear when I make a new thread about the issue. I have been meaning to make a post detailing the findings, but life is really busy for me right now. I plan on doing it tomorrow though. It can get tasks with OPN1 set to max_concurrent = -1. Yes, you get them, but they don't run. Again, this is a dirty trick to circumvent the bug that you have found (in the sense that no CPU-tasks run if you don't want that). There is no way around to have them downloaded, though, simply because WCG appears to consider GPU-tasks related to 'research to run on CPU'. Yes, 0 = no limit. This is why in the profile project limits section it goes: Unlimited, 1, 2, 3 etc. This is similar to some other settings such as "Suspend work if CPU usage is above" X "% of cpu" where "If you set this value to 0, it means no restriction is enforced." Beware that the max_concurrent parameter in app_config.xml does not relate to the number of downloaded/in-queue tasks, but to the number or concurrently running tasks. Again, it's the only parameter that I found that can completely stop running any task. Agreed, I only ran it like that as part of the investigation into the issue. I have since resumed computing the cpu units so they will be completed before the deadline. It would be good for this issue to be fixed though. |
||
|
|
|