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: I
)
I just switched the Moonshot-W machine to 4x threads after running at 3 threads most of today. The cpu seems to not look overloaded with the 12 GPU threads and 4 CPU threads.
The copy graph keeps climbing.
Since I now have a 3 thread baseline from Moonshot-W I will go ahead and bump TheMonolith-W to 3.
I have also toggled the profile for Moonshot-W to no longer run non-preferred tasks in a bid to get some GR tasks back on the GPU. I am going to leave TheMonolith profile allowing non-preferred tasks and see how things work "automagically".
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!
Gary Roberts wrote: Tom M
)
Gary,
I am going to try out the 4 threads for GW gpu on one box while I run 3 on the other.
Now to get some GR tasks :)
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!
must be down to differences
)
must be down to differences in the AMD app vs the nvidia app.
I just did a re-test to re-confirm my earlier tests. running multiples on nvidia cards causes runtimes slower than the multiple, and overall worse production.
on a 2080ti power limited to 225W:
on a 3070 power limited to 200W:
the drop in power consumption is a clear indication that it's doing less work in 2x (hence the slower than half runtimes), despite the gpu utilization being maxed.
_________________________________________________________________________
About the same here, Ian &
)
About the same here, Ian & Steve. Running a GTX1070 at fixed clock & (moderate) voltage with a memory OC. GPU utilization with 1 task is ~80%, with dips to 50-60%. Memory controller load is 70-80%. Going to 2 concurrent WUs increases GPU utilization to >95%, increases the memory controller load a bit and increases the GPU power consumption by ~10%. However, the run times are at best about double, if not a bit worse. If fluctuates a lot and I quickly switched back before getting accurate numbers.
BTW: the O2MDF tasks cause my card to "coil whine" like crazy, which is a sign for heavy load fluctuation. It's better at 2 concurrent WUs (more continous load, I guess), but not worth the price (higher power, loosing 1 CPU thread, questionable GPU performance).
In order to combat coil whine I reduced the GPU voltage & clock strongly and it somewhat helps. Interestingly it hardly affects GPU performance at all. The comparison is a bit difficult due to the fluctuation of runtimes, but it definitely lands in the same ballpark, a few % worse at maximum:
1.860 GHz @ 0.9 V, 95 W: 11:54, 11:25, 11:32
1.518 GHz @ 0.73 V, 73 W: 11:49, 11:44
Please excuse if this has ben reported before, I just read the last page.
MrS
Scanning for our furry friends since Jan 2002
interesting info about the
)
interesting info about the coil whine.
mine exhibit a bunch of whine too (sounds more like crickets lol), but I just accept it. I can't hear it unless I'm right next to the system and I don't sleep in the same room so it doesn't bother me.
_________________________________________________________________________
I'm seeing very low GPU
)
I'm seeing very low GPU utilization on O2MDF tasks on macOS 10.15 (Catalina), and the tasks are taking far more CPU time than I would expect. Tasks take approximately 55min to complete on RX580 (stock clocking) and Intel Core i7-9700K (stock clocking) (https://einsteinathome.org/host/12838735).
how many tasks are you trying
)
how many tasks are you trying to run at once?
on average, the GW GPU tasks require about 2GB VRAM, each. So if you’re running more than about 4 at at time, you will exceed the GPU’s 8GB of memory. From what others have reported with AMD GPUs, this tends to cause it to overflow to system memory and causing massive slowdowns.
if it’s not caused by VRAM overload, then maybe something wrong with the macos app, that GPU utilization is really low. Only 16% explains the slow time.
_________________________________________________________________________
Currently I'm running only 1
)
Currently I'm running only 1 task per GPU (default settings from the project) The system has two RX580s each with 8GB which is why the screenshot shows two tasks. I saw essentially 100% GPU loading on Gamma-ray pulsar binary search tasks. And other projects get 100% GPU loading (PrimeGrid, Collatz). This sure seems like an issue with the macOS O2MDF application. Is there a way I can help debug this?
Justin Gregg
)
I haven't ran the GW GPU apps in a long time; but when they first were released they didn't offload enough of the work from the CPU to the GPU to come close to fully utilizing it, resulting in low numbers because the GPU spent much if its time waiting for the CPU to do stuff. Depending on how much work they've done in the app since then not being able to fully load a GPU isn't a problem with your system but with how the app was written.
seems the GPU GW work is
)
seems the GPU GW work is running thin, and the O2MDF work generator is Disabled. several of my systems are out and mostly getting no work for work requests.
_________________________________________________________________________