Project Resource split not right
22 Apr 2020 17:42:51 UTC
Topic 222091
(moderation:
I have been running Rosetta and Einstein together for the last several years together just fine and have just run into this problem in the last month. I run Einstein on my GPU's and Rosetta on my CPU. It seems that Rosetta is taking all the CPU resource and pushes Einstein down to just one of 4 available slots for my gpu's (2 per GPU). I have not changed any settings, the only thing I have done is paused Einstein about a month ago to work on folding@home.
I have:
-Restarted Boinc
-Changed the resource allocation in my Rosetta account.
-Reset Einstein.
-Tried started one project, then the other.
None of this has helped! Does anyone have any ideas?
Thanks,
-Paul
Did you change anything like
)
Did you change anything like upgrading the Boinc client? Do you know how to create an 'app_config.xml' to control a project. It can be used to set max concurrent tasks.
Yes, I did update the client,
)
Yes, I did update the client, I believe just a few days ago. I am not 100% with the 'app_config.xml', do you have a template?
th3tricky wrote:It seems that
)
Your computers are hidden so nobody can see your hardware details or the particular searches you are contributing to. It's impossible for anyone to be able to know what the problem really is. You talk about "4 available slots" - does that mean you have 4 GPUs installed? I certainly don't imagine so but you need to properly describe things.
You mention "2 per GPU" so I assume you want to run 2 GPU tasks per GPU - which could be fine - if we had any clue about what your GPUs were (hardware wise), how much VRAM, what GPU search(es) you were trying to run, how much CPU support you were supplying to GPU tasks, etc, etc.
Please set out exactly what you would like to achieve and the precise details of all your hardware. Allowing your computers to be visible would be a good first start.
You mention a "template" for an app_config.xml file. Please realise there is no such thing as a "one size fits all" template. If you have been running successfully without such a file, you still should be able to do so. That decision can't be made until the hardware being used and the details of the actual problem are clearly stated. There's no need to over-complicate things if it's not really needed.
Cheers,
Gary.
Fixed. Sorry, that is not
)
Fixed. Sorry, that is not usually something I have hidden.
I have preferences set to run 2 concurrent Einstein tasks per GPU, on 7 cards in 4 separate machines. I have been running that, along with Rosetta on each CPU for over a year with no issues, never having had to mess with resource share settings. As of last week I tried taking my GPU's back to Einstein from folding@home, and Einstein works fine running two tasks per GPU until I enable Rosetta, then it reverts to only one Einstein GPU task running per computer. I do not do Einstein CPU tasks either, just GPU. The only thing I can think of is the recent BOINC update screwed something up.
For GPU's I'm running: RTX2080ti with GTX 1660, RTX2070 with RTX2060, single RTX2070, and 2x GTX 1660's.
Hopefully that helps explain my issue better, if not, please let me know! I'd much rather using my computing power for Einstein.
Put this "app_config.xml"
)
Put this "app_config.xml" file in your rosetti@home project directory (the one under the hidden ProgramData/Boinc folder). Set the # to the number of threads that you used to have rosetti@home running at. This will probably free up the resources to allow you to run E@H on all of your GPUs again.
<app_config>
<project_max_concurrent>6</project_max_concurrent>
</app_config>
Then either cause the BOINC Manager to "read config" files or restart BOINC Manager/Boinc clients.
The total # of R@H threads will reduce to the # you have set and all the GPUs should start running.
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!
th3tricky wrote:Hopefully
)
Yes, indeed! Thanks very much for making your computers visible! :-).
Tom has given you very good advice about limiting the number of threads that Rosetta can consume. That is most likely the reason that BOINC is being restricted with GPU tasks. I would just like to explain why you will probably need to adjust the example supplied for each separate case, if you want to optimise things for the best output. If you want to see the documentation, it is covered here.
For nvidia GPUs, the accepted wisdom is that you need more than a single thread of CPU support for each GPU task, probably at least 1.5 threads - perhaps more. You will need to find out by experiment. I'm guessing that the max concurrent of 6 would probably work well for a 12 thread machine (4x1.5=6) - and even more-so for higher thread counts - but could be a waste of Rosetta CPU output for a 16 thread machine.
Of course, until you find out for sure, it may well be that the opposite could be the case. On the 12 thread machine, your 4 GPU tasks might take a bit longer than they should and you might get better GPU output if you reduced Rosetta to five, leaving 7 threads available for GPU support.
As Tom has already explained, you don't need to stop BOINC to make adjustments. You create or subsequently edit the value in the file whenever you need to and then select "reread config files" in BOINC Manager to get the new value to be immediately accepted.
If you would like to optimise, you need to be aware that the GW GPU tasks can vary quite a bit in crunch time. Once a change is made you should let the new conditions run for a very long time to be certain that the average time has changed in the way you are hoping. Don't be in too much of a hurry :-).
Cheers,
Gary.
Excellent, thanks guys, I
)
Excellent, thanks guys, I will try the config file! When I was running before, next to each task always said .5 CPU, so I assume 4 tasks was taking up 2 threads from the CPU and that setup ran fine. Like you said, I won't know if that was a bottleneck until I give it some more CPU.
Thanks again,
-Paul