GRPS app on iGPU: GPU load
21 Apr 2020 7:50:02 UTC
Topic 222067
(moderation:
I'm running iGPU version (OpenCL) of GRPS app on fairly low-end device, but still only 93% of average GPU load.
It takes full CPU core (by default) too feed GPU still. In your experience - what the reason of decreased GPU load for this particular app?
I could imagine few of them with different possibilities and paths for further optimization.
1. This GPU app behaves same (underloaded GPU) on different GPUs/CPUs combos, it just has too many serial parts. Then One could try running few tasks per device.
2. On some GPU/CPU combos there is ~100% GPU load. So GPU under-feeded by my particular CPU.
Also, what regarding how this particular app's GPU load depends on core reservation? How much load degrades if full core reservation removed?
And finally - what tweaking possibilities exist for this and binary pulsars search GPU app exists if any?
EDIT: it averaged 77% of GPU load after few hours run
I have found the iGPU of
)
I have found the iGPU of Intel current/previous gens to slow the total production of the system down if it is used for crunching.
I see you have two different systems with iGPU's. Which one are you running the "question" system on?
I have just experienced running GW gpu on some Nvidia video cards. The load was 38%-68% per gpu card with 1 cpu driving it. When I tried using 0.5 cpu's per card the gpu load went even lower.
The pulsar gpu app keeps the video cards loaded at the 98% level for the most part.
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!
I'm experimenting with this
)
I'm experimenting with this one currently: https://einsteinathome.org/ru/host/12818330
Have you looked at memory
)
Have you looked at memory usage of both RAM and the video card? Isn't RAM and integrated video sharing memory on most computers?
Jonathan wrote:Have you
)
Just at the system he is talking about. So it is a straight iGPU from Intel. Whole system has 4GB of memory and he is sharing that with the cpu. So life is a bit cramped there.
Tom M
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!
Jonathan написал:Have you
)
Memory usage ~800MB. GPU-Z references it as "dedicated" but of course it's part of system memory.
Same GPU-Z reports sysem memory usage (whole) about 2,4GB.
So, app fits in memory and as long as it fits why should I bother?
I was thinking it was leading
)
I was thinking it was leading to memory swapping and the swapping leading to lower loads.
I see. But project configured
)
I see. But project configured to disallow GPU work for cards with less than 1GB of memory. That's the reason my NV 460SE can't get any work on another host.
So, GPU memory swapping can be ruled out.