My system is running a mix of universe at home and small # of wcg CPU tasks. My GPU is at 4x with the production apparently rising. I will be trying 5x next.
I am assuming a stepwise progression of, in your case, 5, 6, 7, 8 etc.
Your results should be very interesting! At least to me.
Just bumped mine from 4x to 5x and it hasn't crashed (yet). User memory is at now 98% from 79%.
====edit=== Had to take gpu usage down to 0.19 instead of 0.20 before it bumped.
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!
Just bumped mine from 4x to 5x and it hasn't crashed (yet). User memory is at now 98% from 79%.
====edit=== Had to take gpu usage down to 0.19 instead of 0.20 before it bumped.
I will let it run longer. But the preliminary results look like 4x produces more production than 5x. So the only question is was it because the GPU memory was running at 98% or because I have reached that "crossover" point where it does reduce production.
Oh, well, it would have been nice.... but I expect a similar result on other gpus.
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!
While you're in 'replying to questions' mode, here's one that would be very helpful if it could be answered in the affirmative :-).
Since the doubling of credit for current GW, I've noticed a very significant increase in the number of resends arriving on any particular host of mine that I choose to look at. My guess is that there is a new wave of lesser GPUs that aren't coping with the GW tasks. Whilst the h1_... and l1_... data files are <sticky>, the very large skygrid files for each particular 'frequency bin' are not.
There's no problem when a particular bin is still 'current'. As a host progresses from one bin to the next, there will be frequent resends coming for the previous bin, probably for a few weeks until all tasks for that previous bin are completed. I've noticed, on quite a few hosts already, multiple times, that a new resend arrives for a previous bin and the skygrid gets downloaded afresh as well.
So my question is whether or not the skygrid file could be made <sticky>, at least for a few weeks after the primary tasks for a particular bin have all been issued. I've actually started planning for a script to cache all skygrids (I've got about 50 already) and to have each host contribute to and draw from that cache. The plan is for each host to keep track of its skygrids and to immediately replace any that get removed.
It would save me quite a bit of redundant effort if the skygrids were <sticky> until a bin was completed. Would that be possible?
Just bumped mine from 4x to 5x and it hasn't crashed (yet). User memory is at now 98% from 79%.
====edit=== Had to take gpu usage down to 0.19 instead of 0.20 before it bumped.
I will let it run longer. But the preliminary results look like 4x produces more production than 5x.
Spoke too soon. The processing time has resumed running below 3,000 seconds. The first results had that 3,000+ problem. So I will perform a nearly impossible act (sitting on my hands, not changing anything) :)
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!
Just a quick note to follow up on my previous post: I have spread out the 4 GPU tasks I'm running simultaneously over the two chiplets of my 5900X CPU, so they have more CPU cache available, by setting the affinity with Process Lasso, and immediately I'm seeing shorter runtimes. The total board power and utilization of the GPU have improved as well. This is with 4 other CPU tasks running at the same time from another project.
So preliminary testing seems to point to this app loving CPU cache. I wonder whether people with X3D CPUs see substantially shorter runtimes.
I decided to give Process Lasso a try on my Threadripper 2950x. There can be no doubt that confining your various workloads to individual chiplets within the CPU has an immediate positive effect. While I haven't taken the time to quantify the improvement, it was significant enough that I went ahead and purchased Process Lasso!
Thanks for mentioning this program, I had long since forgotten about it.
I seem to only get Meerkat work, no GW work. I tried turning that off and leaving everything else on, but I never get any GW work. BOINC log tells me check E@H log for my machine:
Some older versions of BOINC unreported graphics ram. An updated install my fix the problem.
the issue with older BOINC clients was that they wouldn’t report anything greater than 4GB. But they reported anything less than 4GB fine. Since his client is reporting 2GB, he’s not affected by that issue. he’s on the latest client anyway.
Tom M wrote:Thank you.My
)
Just bumped mine from 4x to 5x and it hasn't crashed (yet). User memory is at now 98% from 79%.
====edit=== Had to take gpu usage down to 0.19 instead of 0.20 before it bumped.
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: Just bumped
)
I will let it run longer. But the preliminary results look like 4x produces more production than 5x. So the only question is was it because the GPU memory was running at 98% or because I have reached that "crossover" point where it does reduce production.
Oh, well, it would have been nice.... but I expect a similar result on other gpus.
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!
Bikeman (Heinz-Bernd
)
While you're in 'replying to questions' mode, here's one that would be very helpful if it could be answered in the affirmative :-).
Since the doubling of credit for current GW, I've noticed a very significant increase in the number of resends arriving on any particular host of mine that I choose to look at. My guess is that there is a new wave of lesser GPUs that aren't coping with the GW tasks. Whilst the h1_... and l1_... data files are <sticky>, the very large skygrid files for each particular 'frequency bin' are not.
There's no problem when a particular bin is still 'current'. As a host progresses from one bin to the next, there will be frequent resends coming for the previous bin, probably for a few weeks until all tasks for that previous bin are completed. I've noticed, on quite a few hosts already, multiple times, that a new resend arrives for a previous bin and the skygrid gets downloaded afresh as well.
So my question is whether or not the skygrid file could be made <sticky>, at least for a few weeks after the primary tasks for a particular bin have all been issued. I've actually started planning for a script to cache all skygrids (I've got about 50 already) and to have each host contribute to and draw from that cache. The plan is for each host to keep track of its skygrids and to immediately replace any that get removed.
It would save me quite a bit of redundant effort if the skygrids were <sticky> until a bin was completed. Would that be possible?
Cheers,
Gary.
Tom M wrote: Tom M
)
Spoke too soon. The processing time has resumed running below 3,000 seconds. The first results had that 3,000+ problem. So I will perform a nearly impossible act (sitting on my hands, not changing anything) :)
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!
fastbunny wrote: Just a
)
I decided to give Process Lasso a try on my Threadripper 2950x. There can be no doubt that confining your various workloads to individual chiplets within the CPU has an immediate positive effect. While I haven't taken the time to quantify the improvement, it was significant enough that I went ahead and purchased Process Lasso!
Thanks for mentioning this program, I had long since forgotten about it.
Regards,
G
I seem to only get Meerkat
)
I seem to only get Meerkat work, no GW work. I tried turning that off and leaving everything else on, but I never get any GW work. BOINC log tells me check E@H log for my machine:
1/28/2024 4:20:53 PM | Einstein@Home | see scheduler log messages on https://einsteinathome.org/host/13171349/log
I don't see anything obvious why I can't get any GW work.
The GPU is an 960M with 4mb of VRAM.
boinc identifies
)
boinc identifies your GPU as only having 2GB of VRAM, which is not enough to get the GW tasks, you need at least 3GB
_________________________________________________________________________
Some older versions of BOINC
)
Some older versions of BOINC unreported graphics ram. An updated install my fix the problem.
Ben Scott wrote:Some older
)
the issue with older BOINC clients was that they wouldn’t report anything greater than 4GB. But they reported anything less than 4GB fine. Since his client is reporting 2GB, he’s not affected by that issue. he’s on the latest client anyway.
_________________________________________________________________________
Ian&Steve C. wrote: boinc
)
Ah...I see. I wonder why I was thinking it had 4gb... Must have been thinking of something else...wishful thinking? ^_^