I just enabled the "GPU utilization factor of BRP apps" in the Apple Silicon plan classes, i.e. you should be able to experiment with running multiple GPU tasks in parallel (see Account -> Preferences -> Project).
This may help with getting the GPU tasks finished faster.
Hey BM - thank you and your team for all that you do!
I was trying to take advantage of multiple GPU tasks, but I've had no luck. Last week I changed my utilization preference for "GPU utilization factor of BRP apps" to 0.50, but I don't see any differences. I still just have one BRP4A task running. Any suggestions?
I was trying to take advantage of multiple GPU tasks, but I've had no luck. Last week I changed my utilization preference for "GPU utilization factor of BRP apps" to 0.50, but I don't see any differences. I still just have one BRP4A task running. Any suggestions?
Have you checked your preferences? Your compute and project preferences need to be on the same venue, i.e.: School, Work, Home.
Also, do you have an app_config.xml file? You'll need that to acquire more than one task at a time.
I was trying to take advantage of multiple GPU tasks, but I've had no luck. Last week I changed my utilization preference for "GPU utilization factor of BRP apps" to 0.50, but I don't see any differences. I still just have one BRP4A task running. Any suggestions?
I am having the same issue ….. changed my GPU Utilization Factor for BRP apps to [0.5] but it is still crunching only one GPU task at a time. Such a change has always worked immediately on Linux, MacOS and Windows so I suspect something is not working correctly in the Apple Silicon code for these work units. This Utilization Factor change change has been in place and not working for several days now. Hopefully Bernd can get someone to take a look.
... Such a change has always worked immediately ...
Just to be perfectly clear, the change can only occur after your boinc client contacts the project and is advised of the setting change. That might happen relatively quickly but you can force it by clicking 'update' in the manager after you save the change on the website.
The two big advantages of GPU utilization factor are not having to construct and deploy the more complicated app_config.xml file and being able to reverse it quickly and easily if you decide not to continue with concurrent tasks. The disadvantage is that you can't change other factors that you might need, such as cpu_usage, max_concurrent, etc.
I suspect something is not working correctly in the Apple Silicon code for these work units.
The "Apple Silicon code" (application) doesn't have anything to do with the task/resource assignments done by the BOINC client. But something is clearly odd here. How many (suitable) pending tasks do you see in the work queue? More than one?
Update: yes, clearly... :)
Update 2: I changed my preference to 0.5 GPU usage for BRP and BOINC 8.02 runs two GPU tasks (and some extra CPU tasks) in parallel. Please make double sure you set your preferences on the web as needed and that you're not overriding them locally in the BOINC manager. For further insight I recommend you enable some extra client logging flags, namely: coproc_debug and maybecpu_sched, cpu_sched_debug and cpu_sched_status as well.
Just to be perfectly clear, the change can only occur after your boinc client contacts the project and is advised of the setting change.
Change will take even longer. Although this parameter is set in the project preferences, it's actually not communicated to the client as a project-wide setting (to be precise it is, but the client doesn't do anything with it). Instead it affects the assignment of (new) tasks to the client on the server side.
Changing this setting results in a different "ngpu" setting of newly assigned tasks. This means than when you change this setting, as long as you have "old" tasks in the cache, you will not notice a difference. Only when you start running tasks that have been assigned tor you after you changed that setting.
Just to be perfectly clear, the change can only occur after your boinc client contacts the project and is advised of the setting change.
Change will take even longer.
Thanks for the extra information and indeed, I was a bit sloppy with my particular wording.
Instead of "boinc client contacts ...", I should have written "boinc client requests new work from ...".
Also, your 2nd paragraph suggests something I was not aware of (or at least I don't remember).
Since there are only 4 locations (venues) available, and I have a large number of quite different host/GPU combinations, I essentially need the ability to control each one individually. I would have tried the web based utilization factor when you first created it many years ago, but quickly moved to individual app_config.xml files for the needed flexibility. Hence, I don't really remember what you mention - ie. that "old tasks" continue with the former multiplicity and that you have to wait for "new tasks" to reach the top of the queue.
With app_config, the change is immediate and I wrongly assumed it would be the same with the GPU utilization factor. Sorry about that.
Sometime in the past 12-14 hours BOINC manager delivered a new message that a new version on BOINC was available, 8.0.2 ….. also mentioned by Oliver above, so it made me look. Downloading and installing that did two things on restarting BOINC, first it automatically deleted all of the old GPU tasks, and second, it downloaded a bunch of new ones. The result is that without doing anything else, BOINC is now running two GPU tasks. So, the key here is updating to 8.0.2, something I don’t recall reading about, but it seems to work just fine. Thanks for all the comments made above this which steered me in the right direction !
No, it's not. By deleting all local tasks the upgrade forced new downloads which in turn did what Bernd just described above. So it's the new tasks, not the new BOINC version. Also, if BOINC really removed existing (in progress) tasks, that's a huge bug, I'd say.
Bernd Machenschalk wrote: I
)
Hey BM - thank you and your team for all that you do!
I was trying to take advantage of multiple GPU tasks, but I've had no luck. Last week I changed my utilization preference for "GPU utilization factor of BRP apps" to 0.50, but I don't see any differences. I still just have one BRP4A task running. Any suggestions?
jd wrote: I was trying to
)
Have you checked your preferences? Your compute and project preferences need to be on the same venue, i.e.: School, Work, Home.
Also, do you have an app_config.xml file? You'll need that to acquire more than one task at a time.
https://boinc.berkeley.edu/wiki/User_manual
Proud member of the Old Farts Association
You don't need to use an
)
You don't need to use an app_config file if you just change your utilization preferences on the website.
jd wrote: I was trying to
)
I am having the same issue ….. changed my GPU Utilization Factor for BRP apps to [0.5] but it is still crunching only one GPU task at a time. Such a change has always worked immediately on Linux, MacOS and Windows so I suspect something is not working correctly in the Apple Silicon code for these work units. This Utilization Factor change change has been in place and not working for several days now. Hopefully Bernd can get someone to take a look.
Richard de Lhorbe wrote: ...
)
Just to be perfectly clear, the change can only occur after your boinc client contacts the project and is advised of the setting change. That might happen relatively quickly but you can force it by clicking 'update' in the manager after you save the change on the website.
The two big advantages of GPU utilization factor are not having to construct and deploy the more complicated app_config.xml file and being able to reverse it quickly and easily if you decide not to continue with concurrent tasks. The disadvantage is that you can't change other factors that you might need, such as cpu_usage, max_concurrent, etc.
Cheers,
Gary.
Richard de Lhorbe wrote:I
)
The "Apple Silicon code" (application) doesn't have anything to do with the task/resource assignments done by the BOINC client. But something is clearly odd here. How many (suitable) pending tasks do you see in the work queue? More than one?
Update: yes, clearly... :)
Update 2: I changed my preference to 0.5 GPU usage for BRP and BOINC 8.02 runs two GPU tasks (and some extra CPU tasks) in parallel. Please make double sure you set your preferences on the web as needed and that you're not overriding them locally in the BOINC manager. For further insight I recommend you enable some extra client logging flags, namely: coproc_debug and maybe cpu_sched, cpu_sched_debug and cpu_sched_status as well.
Hope this helps
Einstein@Home Project
Gary Roberts wrote:Just to
)
Change will take even longer. Although this parameter is set in the project preferences, it's actually not communicated to the client as a project-wide setting (to be precise it is, but the client doesn't do anything with it). Instead it affects the assignment of (new) tasks to the client on the server side.
Changing this setting results in a different "ngpu" setting of newly assigned tasks. This means than when you change this setting, as long as you have "old" tasks in the cache, you will not notice a difference. Only when you start running tasks that have been assigned tor you after you changed that setting.
BM
Bernd Machenschalk
)
Thanks for the extra information and indeed, I was a bit sloppy with my particular wording.
Instead of "boinc client contacts ...", I should have written "boinc client requests new work from ...".
Also, your 2nd paragraph suggests something I was not aware of (or at least I don't remember).
Since there are only 4 locations (venues) available, and I have a large number of quite different host/GPU combinations, I essentially need the ability to control each one individually. I would have tried the web based utilization factor when you first created it many years ago, but quickly moved to individual app_config.xml files for the needed flexibility. Hence, I don't really remember what you mention - ie. that "old tasks" continue with the former multiplicity and that you have to wait for "new tasks" to reach the top of the queue.
With app_config, the change is immediate and I wrongly assumed it would be the same with the GPU utilization factor. Sorry about that.
Cheers,
Gary.
Sometime in the past 12-14
)
Sometime in the past 12-14 hours BOINC manager delivered a new message that a new version on BOINC was available, 8.0.2 ….. also mentioned by Oliver above, so it made me look. Downloading and installing that did two things on restarting BOINC, first it automatically deleted all of the old GPU tasks, and second, it downloaded a bunch of new ones. The result is that without doing anything else, BOINC is now running two GPU tasks. So, the key here is updating to 8.0.2, something I don’t recall reading about, but it seems to work just fine. Thanks for all the comments made above this which steered me in the right direction !
Richard de Lhorbe wrote:So,
)
No, it's not. By deleting all local tasks the upgrade forced new downloads which in turn did what Bernd just described above. So it's the new tasks, not the new BOINC version. Also, if BOINC really removed existing (in progress) tasks, that's a huge bug, I'd say.
Anyway, glad that it works for you now.
Oliver
Einstein@Home Project