How to get bigger tasks for Apple GPU or more than 1088 tasks.
2 Jul 2024 15:42:48 UTC
Topic 231268
(moderation:
Hi there.
I have a system on Apple M2 Pro that can handle 12 parallel GPU tasks (Binary Radio Pulsar Search) in about 9 minutes per task. With that speed of processing, after about 16 hours, it runs out of work to do due to the daily limit of 1088 tasks per host. How to get bigger tasks (with more credits) or more small tasks?
Not sure if I understand your
)
Not sure if I understand your use of the terminology "parallel" for the gpu tasks.
Are you running 9 tasks concurrently at the same time on a single gpu? IOW, on the same device?
Via an app_config.xml entry for gpu_usage=0.1 or via the project preferences gpu concurrency setting for 0.1?
Surprised that 9 tasks will fit in the VRAM storage without complaint or errors.
You might try and use the ncpus option in the cc_config.xml file to spoof the cpu count to tell Boinc that you have more than the 9 cpu cores it detects. That might bump your daily quota allocation up.
That has always been the solution to get more cpu tasks allocated on fast cpu hosts. Don't know if it would apply to the gpu tasks though, I don't know anything about Apple hardware for a better guess.
Yeah, gpu_usage=0.1 (or
)
Yeah, gpu_usage=0.1 (or something about it), 10 tasks in parallel. This SOC can address most of the memory for the CPU and GPU simultaneously. Mine says 10GB of memory is available for the GPU. And as all its memory, HBM is crunching such tasks quite quickly. The question is: how can we utilize it completely?
The machine stats: https://einsteinathome.org/host/13189042
memory utilization is not the
)
memory utilization is not the measure of gpu utilization. the GPU app will be able to use the GPU fully. maybe 2 or 3 tasks at a time to fill in the gaps. certainly 9 or 10 is too many.
the M2 chip does not use HBM. it has normal memory packaged more closely to the CPU cores, but that does not make it "HBM".
you should go back to 1x and check overall tasks/day throughput and recheck as you add 2 or 3 tasks per GPU. but with a small GPU like in the Apple silicon M-chips, I'd imagine that 1x is enough to more or less fully utilize the GPU cores.
_________________________________________________________________________
I see how many tasks are
)
I see how many tasks are reported and confirmed. I calculated the maximum PPD, and yes, it runs 10 tasks in parallel without issues.
When I run a single task, it takes about 5 minutes; when I run ten tasks, it takes about 9 minutes per task. It calculates all of them in parallel without issues. Getting more than 10 tasks is not practical as SOC starts overheating.
If you wish you can check details here https://einsteinathome.org/account/322067/computers
The trick with ncpus worked
)
The trick with ncpus worked out. Thank you!
smile wrote:The trick with
)
Happy that helped. What value did you use to spoof additional cpus? I only see 991 tasks in progress. That won't be enough to avoid your 1088 tasks a day limit.
Really surprised that M2 chip can do so many concurrent gpu tasks without errors.
That is a really efficient app then. Kudos.
I have regularly used ncpus
)
I have regularly used ncpus to spoof the server into sending me more GPU tasks.
I also have to limit all the CPU tasks so the boincmgr doesn't appear "run" more CPU tasks than are presently on the cpu.
Usually 2 to 4 times the actual CPU count were sufficient for the ncpus parameter.
I ran into this problem with a 16 thread Ryzen and 3 GPU's.
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
smile wrote: 12 parallel GPU
)
Out of curiousity, if I calculate that that would lead to a RAC of 120.000 - 150.000.
Is this a mobile or desktop system?
Mobile, 16 inch laptop.so far
)
Mobile, 16 inch laptop.so far i have about 100k daily