On the new 3011L03 tasks for FGRP
15 Dec 2021 3:41:50 UTC
Topic 226591
(moderation:
Recently on my computer I started getting 3011L03 GPU tasks for the Gamma-ray subproject. These tasks seem to run faster, however they behave rather weirdly - the old 4013 tasks would jump from 89.997% to finished instantaneously, but these tasks hang on 89.997% for quite a while. I remember someone saying that this part is supposed to be the CPU checking for matches, but during these three or four minutes hanging at 89% the GPU is at maximum utilization too. Moreover, the GPU power usage drops to 8W for an unknown reason (normally at 23W when running the main part of the tasks) and the Xorg Window Manager also becomes incredibly laggy.
Of course, it might be that my specs are terrible (they are), but the GPU manages the first 89% of the workunit just fine.
What is special about these new workunits and is there a workaround for these three minutes of hanging and lag? Thank you very much!
are you running CPU work
)
are you running CPU work also? if your CPU is pegged at 100% doing other stuff, then I could see why it would take 3 mins for the last part. I'm not sure how involved the GPU itself is during this time. but if the CPU is too busy I could see that holding you back,
for comparison, my system takes about 15 seconds from 89 to 100%, on a 5950X (but its only using 1 core for the task) and a 3080Ti.
I'm running 4 tasks at a time on these 3011-series tasks. the 4000-series tasks do best at 1 at a time, but these ones see a small benefit with multiples. I don't think you'll be able to run multiples on your 750 though with only 2GB VRAM. each task needs about 1GB.
_________________________________________________________________________
I don't have anything to add
)
I don't have anything to add on the end-stage behavior, but as the thread title may attract people with other interests, I'll take the opportunity to post performance comparisons.
One might expect that as all three of my systems are running Navi or Big Navi AMD GPU cards, with Intel CPU hosts, under Windows 10, that the performance shift would be very nearly the same, but it differs more than I'd have guessed.
Stoll9
RX 6800 GPU throttled down to 2000 MHz GPU clock
i5-4690K CPU
3X multiplicity
35% production increase
Stoll10
RX 6800XT GPU throttled down to 2000 MHz GPU clock
i5-9400F CPU
4X multiplicity
40% production increase
Stoll11
RX 5700 GPU throttled down to -40% power limit
i3-10100F CPU
3X multiplicity
27% production increase
The production increase numbers compare recent throughput running tasks of the new LATeah3011L03 flavor to recent tasks of the LATeah4013 flavor.
The changeover for predominant flavor was just about 24 hours ago.
Ian&Steve C. wrote: are you
)
Thank you for your response! My other CPU is currently running WCG, but suspending it didn't help - the final 10% still took about 3 minutes - I guess my CPU is too slow. During the final 10% even though the GPU is supposedly still at 100% utilization the FGRP GPU process only uses 24 MB of memory - maybe it's holding up the GPU?
The peak GPU memory usage of the process is about 730 MB, so I might be able to do two at the same time, but I only have two cores and I'm afraid it might overload and damage my GPU.
Bumping this thread because
)
Bumping this thread because there seems like no viable fix. Would it be possible to only accept 4103 tasks? The final 10% take an enormous load on my computer (relative to the rest of the work unit) and lags the display.
The last 10% of the task
)
The last 10% of the task actually is running on the cpu, not the gpu so it shouldn't be lagging your display.
The system memory consumption goes up during that last 10% so you should either add more system memory or reduce the number of cpu tasks running.
Always good computer discipline to not use 100% of the cpu for BOINC. You should reserve 1-2 cores of the cpu for system maintenance and responsiveness. You can set that in the Boinc Manager Options >> Computing Preferences. 90-95% of the cpu is usually a good number to try and leaves a few cores free for the OS.
the last 10% does appear to
)
the last 10% does appear to be using the GPU. I think it's doing some stuff in double precision in conjunction with CPU processing. very weak DP cards will take longer. even though my "last 10%" portion only takes tens of seconds, I can see some GPU utilization still happening during this time. 10-50%, but power used drops a lot.
_________________________________________________________________________
Yep, the GPU load goes up and
)
Yep, the GPU load goes up and down during that final stage and I see lagging on screen (GTX970) if I am watching video at the same time. This lasts for about 45 seconds.
The server problem yesterday
)
The server problem yesterday has caused some download errors and a lot of resends for these tasks. I've got one task that is the 13th copy of the task https://einsteinathome.org/workunit/594930364