Gravity Wave tasks that run on a cpu only system?
26 Nov 2023 14:49:55 UTC
Topic 230411
(moderation:
I get the impression that Gravity Wave research is the center of research here but the Pulsar searches are here to keep the volunteers engaged between GW research projects.
One of the systems I am running is a CPU-only system. Are there ANY CPU-only GW tasks? How should I be setting up my profile to access them?
Thank ;you.
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!
there used to be CPU only GW
)
there used to be CPU only GW search in previous iterations, they always ended in '1' on the task type. like O2MD1, O3MD1, etc. and the GPU searches usually ended in 'F'
however, for the current search, it doesn't follow that convention. there is only a GPU search for the current O3AS tasks.
as I understand it, the BRP searches are not simply for engagement, but to find likely targets and calibration sources for the GW search. both are useful science for this project.
_________________________________________________________________________
Ian&Steve C. wrote: as I
)
I stand corrected. I guess even a gtx 1060 6GB that can run allsky would be better than nothing then.
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!
I remember crunching O3MD1,
)
I remember crunching O3MD1, it used quite a lot of RAM, I even upgraded to a 64Gb kit, but now is pretty much idling, using just a small portion, only crunching CPU tasks.
The only cpu only
)
The only cpu only sub-projects left now are the Gamma Ray Pulsar Search#5 and the Binary Pulsar Search Long tasks.
And both of them are VERY light in RAM usage.
Keith Myers wrote: The only
)
A couple of interesting observations from my Epyc-7742-ES-QS system. This is running with a 126 task limit on e@h. And a mix of the above two sub-projects.
I have just disabled the "keep non-gpu tasks in memory" to see if the total memory usage declines over time. My guess is that some projects run small enough tasks that most of the memory accesses are from the CPU cache. And that these e@h tasks are not like that.
This is not the first time I have run into this kind of bottle-neck. I believe I had the same issue with Rosetti@home CPU tasks on a Ryzen 3950x CPU. Above 4-6 CPU tasks there was a massive processing slowdown.
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!
is that the system with the
)
is that the system with the broken memory channel?
might want to add more ram to fill the channels and get more bandwidth. more tasks = more required bandwidth usually.
_________________________________________________________________________
Ian&Steve C. wrote: is that
)
Yes it has a broken memory channel. I could add 3 more ram chips. I have tendency to like round numbers in ram chips. So the question would be 6 or 7 memory chips be a better choice?
After I turned off the "keep non-gpu tasks in memory" I rebooted. Used memory went down to about 50 percent and virtual memory went from about 10 percent to nearly zero.
Not having anything paging out to the m2 ssd might be a major difference.
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!
Tom M wrote: Not having
)
I think that not keeping the tasks in memory added another 30-45% to the processing time.
I have turned that back on.
I will get the box opened up and pull the gpu as well as add ram possibly tomorrow.
May end up running at 50% of available threads or something.
Just as long as the heat it is throwing off doesn't go down too much.
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!
Let me ask something...
)
Let me ask something... please, correct me if I'm wrong, as far as I understand, the GPU GW(O3AS) tasks depends a lot on the CPU performance, when I dedicate the CPU only for the GPU tasks, I got much better completion times, if I try to crunch CPU-only tasks at the same time, the O3AS tasks running on the GPU takes much longer. My CPU is quite old to work together with my new RX 7800 XT. (13139360)(not installed right now), the GPU stays idle quite a lot, even when running 2 or 3 tasks at the same time. If I disable Hyperthreading (or SMT on AMD) and use a full core, instead of only a thread, will I get better completion times?
Rodrigo wrote:Let me ask
)
===edit====! Some how I missed you have hyper-threading. mumble...
YMMV (Your Milage may vary) but it sounds like running mostly gpu tasks is the way to go since they "pay off" more and use less time.
Your running AMD Ryzen 3 4100 4-Core with a Radeon RX 580 (I think).
===edit== Try setting your CPU to 87.5% cores/100% CPU time. Then set your AllSky to a single task. No CPU tasks. This will give you 7 CPU cores to play with.
How long are they taking to process? That is your baseline. I would run at least 10 tasks to make sure you have a good idea of the average processing time.
My AllSky single tasking on a Nvidia Gtx 1060 6 GB is here. ===edit=== Mine defaults to 0.9 cpus per gpu task.
After you have your baseline, try 2x tasks. If the time it now takes to process is LESS THAN DOUBLE the baseline you have a net gain and should be running at least 2x. Ditto on 3x. Is the time less than triple of the baseline?
Once you have figured out if running less than 3 GPU tasks is your most productive then any of the CPU threads not in use can be used for CPU crunching. ===edit== This means you can run upto 6 CPU tasks.
Most of the time you will want to run a NON-einstein at home cpu task on your cpu. e@h does not mix cpu and gpu tasks well due to the way the scheduler seems to work.
I like PrimeGrid, World Community Grid and others for my alternate CPU tasks. Universe at Home runs somewhat "slowly" under Windows compared to Linux so you may not want to run it on your machine(s).
===edit==
You can apply the same logic on your Ryzen 3600. Not sure how fast those Radeon HD 7700 Series will process or even if they have enough ram to run any e@h gpu tasks. ===edit== I see you are running PrimeGrid tasks. Maybe they are gpu tasks?
HTH,
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!