Thank you for the summary. Since my goal was to speed up processing by speeding up what appeared to be a CPU bottleneck I appreciate it.
I have switched to the GPU-only version.
I am hoping for a general trend of lower task times. So far the main thing I am noticing is less variability in the task times.
I have also been reading Keith and your discussion. Thank you both.
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!
Since my goal was to speed up processing by speeding up what appeared to be a CPU bottleneck I appreciate it.
I add my observation to this since I use AMD cards there is no cuda for me.
My host with the Intel i7 10750 uses much more CPU time than my my machine with an AMD 7800 X3d per task.
Precisely it's 1200 seconds for the Intel versus 300 seconds for the AMD CPU. Yet besides the CPU only part at the beginning, end and in the middle this has no effect on the computing time. It just uses more CPU time, that's it.
However, when I use the CPU for CPU tasks in addition and I use more than half of the cores minus 1 then GPU calculation slows down. But as long as there 1 free core per GPU task the system will not make use of hyperthreading for any of the Einstein tasks and not slow down you GPU.
I just switched over to the 1.14 app on my 7713 Eypc host. Will see if I have the experience as Ian.
The 1.14 app is definitely more consistent on the Ryzen hosts. And after a day on the 5950X host I gained shorter crunching times. So a win there. I also just switched over the 7950X hosts and hope to have the same gain as the 5950X host.
[Edit]
I am getting almost a 25% reduction in crunching time for the 1.14 app over the older 1.08 app on my 7950X hosts now.
But not nearly as much on the 7713 host. Maybe 5%? But I do see more consistency in all hosts for crunching time and not as much variability. Think that there being more contention for cpu resources on the Epyc and its core clocks speeds compared to the rather locked in full boost speed on the 7950X hosts has most to do with the difference in improvement.
This "chunky" release seem to confuse not only the scheduler, but us as well. We need t think about and work on this a little more. For now I will disable "locality scheduling" and the O3ASHF search.
That's a shame now that I am getting such good results from the 1.14 app. Oh well . . .
Brp7/MeerKat "forever" :)
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!
You don't use any of your CPU's at Einstein? just curious.
Paste experience with the project shows that it doesn't intermix very well on GPU and CPU tasks on the same computer due to cache/buffer behavior. If you add CPU apps to a system with lots of GPU "experience" there is a tendency to download nothing but CPU tasks. I am not going to try to explain why because I am a bit hazy on the technical details. But...
There are two workarounds/fixes. One involves reducing the task cache to 0.01 days. And manually switching profiles depending on which type of task is "starving".
The other is using a custom boincmgr client like Pandora (I don't believe it is publicly available). If you are running Windows only the first workaround/fix is available.
Respectfully,
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!
The daily driver I am testing
)
The daily driver I am testing 1.14 app on has only DDR4 and a 5950X locked at 4.4 Ghz with the memory running at 3600Mhz and CL14 tight timings.
The cards are a 3090 and 3080 Ti. After a lot more tasks done, the times are basically the same as the 1.08 app.
Ian& SteveC, Thank you for
)
Ian& SteveC,
Thank you for the summary. Since my goal was to speed up processing by speeding up what appeared to be a CPU bottleneck I appreciate it.
I have switched to the GPU-only version.
I am hoping for a general trend of lower task times. So far the main thing I am noticing is less variability in the task times.
I have also been reading Keith and your discussion. Thank you both.
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!
Tom M wrote: Since my goal
)
I add my observation to this since I use AMD cards there is no cuda for me.
My host with the Intel i7 10750 uses much more CPU time than my my machine with an AMD 7800 X3d per task.
Precisely it's 1200 seconds for the Intel versus 300 seconds for the AMD CPU. Yet besides the CPU only part at the beginning, end and in the middle this has no effect on the computing time. It just uses more CPU time, that's it.
However, when I use the CPU for CPU tasks in addition and I use more than half of the cores minus 1 then GPU calculation slows down. But as long as there 1 free core per GPU task the system will not make use of hyperthreading for any of the Einstein tasks and not slow down you GPU.
I just switched over to the
)
I just switched over to the 1.14 app on my 7713 Eypc host. Will see if I have the experience as Ian.
The 1.14 app is definitely more consistent on the Ryzen hosts. And after a day on the 5950X host I gained shorter crunching times. So a win there. I also just switched over the 7950X hosts and hope to have the same gain as the 5950X host.
[Edit]
I am getting almost a 25% reduction in crunching time for the 1.14 app over the older 1.08 app on my 7950X hosts now.
But not nearly as much on the 7713 host. Maybe 5%? But I do see more consistency in all hosts for crunching time and not as much variability. Think that there being more contention for cpu resources on the Epyc and its core clocks speeds compared to the rather locked in full boost speed on the 7950X hosts has most to do with the difference in improvement.
FYI, O3AS is disabled for
)
FYI, O3AS is disabled for now. there are tasks in the queue, but they are not being sent out.
https://einsteinathome.org/content/all-sky-gravitational-wave-search-o3-data-o3ashf1?page=5#comment-223946
_________________________________________________________________________
That's a shame now that I am
)
That's a shame now that I am getting such good results from the 1.14 app. Oh well . . .
Keith Myers wrote: That's a
)
Brp7/MeerKat "forever" :)
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!
Hi Keith, You don't use
)
Hi Keith,
You don't use any of your CPU's at Einstein? just curious.
I generally don't use cpus
)
I generally don't use cpus when gpu apps are available and vice versa.
But with the 1.08 app I used plenty of the cpus at the mid and endpoints for the O3AS gpu tasks.
Currently the cpus on my hosts are doing the Boinc Games Sprint on Milkyway cpu N-body tasks.
Normal conditions the cpus are doing Asteroids and yoyo.
Filipe wrote: Hi Keith, You
)
Paste experience with the project shows that it doesn't intermix very well on GPU and CPU tasks on the same computer due to cache/buffer behavior. If you add CPU apps to a system with lots of GPU "experience" there is a tendency to download nothing but CPU tasks. I am not going to try to explain why because I am a bit hazy on the technical details. But...
There are two workarounds/fixes. One involves reducing the task cache to 0.01 days. And manually switching profiles depending on which type of task is "starving".
The other is using a custom boincmgr client like Pandora (I don't believe it is publicly available). If you are running Windows only the first workaround/fix is available.
Respectfully,
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!