Discussion Thread for the Continuous GW Search known as O2MD1 (now O2MDF - GPUs only)
Language
Copyright © 2024 Einstein@Home. All rights reserved.
Copyright © 2024 Einstein@Home. All rights reserved.
Ian&Steve C. wrote:is there a
)
For anyone reading here and wondering about this, the 2.08 and 2.08-beta are actually the same app so there is no point running the beta app and expecting to see any difference.
A couple of weeks ago, I pointed out the reason for the beta app in this message. It was just to solve a problem if using ROCm to supply the OpenCL compute capability. If you check this link, you will find further links to the full story. The beta app was just the solution to the problem and there were no performance mods that came with it. Bernd announced the beta app on 21st Feb and then promoted it to be the standard app on 24th Feb - the very last message in the full original thread.
As a general piece of advice, unless you are paying really close attention to what is going on, it's probably safer not to have test apps set. It's quite easy to make the change as soon as a new test app is seen and verified to be working sensibly and not immediately causing all tasks to fail.
Cheers,
Gary.
Gary Roberts wrote: As a
)
I don't closely monitor things all the time, but I see no problem with me running test apps. Presumably when they fail, someone at the other end sees why they failed, so it's helping the development of the program. If a programmer wants more information, they're free to email me for logs. I have the email me box checked too.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Ian&Steve C. wrote:is there
)
I couldn't get the production version to run more than 4 copies. Apparently my latest attempt to run 6 tasks has failed. I can't seem to run 5 tasks. It reverted to 4 tasks when I set it to 0.20 gpu usage.
I have just switched off the beta toggle. Will try higher thread counts after it clears the beta tasks.
Apparently I can increment by 2's but not odd #'s which is extremely odd. Which also may be where I got the "can't run more than 4 tasks" result because I was trying for 5.
"More testing is needed" to quote many other people. :)
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
Tom M wrote: I couldn't get
)
I've run 3 before no problem, I'm not aware of an odd number limitation. But that was a while ago, so the coding may have changed since then.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.
Tom M wrote:I can't seem to
)
I seem to recall someone else a while ago mentioning that to get 5 tasks, the value needed to be just below 0.20. It might be worth trying 0.19 to see if that gives you 5 running.
Cheers,
Gary.
Gary Roberts wrote: Tom M
)
I have set it down to 0.166 and then used the maximum_tasks to limit it to 5 GW tasks which seems to be the GW gpu limit. The gpu tasks stall if I try to run 6 at the same time.
And that also has gotten me to wondering. Does anyone have an opinion on the best total production for GW gpu on a Radeon VII?
Would running two tasks (for instance) increase my total production?
I currently think that running 5 simultaneous gpu tasks produces more total production. I probably need to re-test but it would be nice if someone else has data points.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
With my 3570k and Linux Mint
)
With my 3570k and Linux Mint I run up to 4 WUs on the VII from time to time, the 3700x is faster and can also process 5 WUs (with factor 0.19).
The 3570k is loaded with 4 WUs, the 3700x could probably do more than 5.
Driver 20.10 for Ubuntu, Powerlimit @ 180 W, 1000 mV core voltage.
Tom M wrote:I I probably need
)
Earlier today (on the Radeon VII box) I set both the GW gpu and Gamma Ray Pulsar#1 gpu tasks to 1:1 to get more baseline data points.
The RAC has been collapsing anyway so I no longer have any production RAC's to "protect".
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor)
A few observations on this
)
A few observations on this data run.
It is well known that > 3GB of memory is required on Nvidea cards.
I have a GTX1660Super which has 6GB and Ian advised me to only run 1X rather than 2X.
I tried 1X and all was well as expected. Then I went back to 2 X and monitored memory usage very closely for some days. A data set starts out for several runs using a bit over 2GB then it switches over to using 4GB and lastly it starts using just a bit under 6GB. There have been no errors so far. I am showing 437 valid tasks with no errors or invalids.
The card will process 3 an hour running 1X or 4 an hour running 2X.
I speculate that when running multiple units there might be some shared memory units because the 3GB GTX1060s I have bomb out 1/3 of the @ 1X which is consistent with my memory monitoring if some of the memory is shared.
Betreger wrote: A few
)
I guess either:
Your 6GB card is better designed and can handle some memory being system RAM.
The chances of having two large WUs at once are low, so the total never exceeds 6GB.
If this page takes an hour to load, reduce posts per page to 20 in your settings, then the tinpot 486 Einstein uses can handle it.