Einstein BRP7 Linux/Nvidia Special app
20 Oct 2023 22:14:42 UTC
Topic 230230
(moderation:
Here's another special optimized app from Petri, this time for BRP7 :) We've actually had this for quite a while, but there wasn't much interest since FGRPB1G was still going and most people were running that. plus there was some strangeness with validation that seems to have been finally sorted out.
from mine and others' testing, running 1x tasks per GPU should produce the best production, but feel free to try out multiples for yourself.
VRAM use is pretty low. and CPU use is also pretty low, about 5-10% CPU core utilization.
expect 30-40% speed boost over the stock nvidia app depending on your exact configuration.
Requirements:
- This requires Linux. there will not be a Windows version.
- This requires and is only applicable for modern Nvidia GPUs. Maxwell through Ada (GTX900 - RTX4000 series)
- It's compiled for CUDA. there will not be an AMD version
- This requires the CUDA 12.x drivers, need at least the 525 branch or later.
- This requires GLIBC 2.27 or better (Ubuntu 18.04 and up)
Get it here: https://drive.google.com/drive/folders/1UlXGRNuEydmVT_hfge_XNVMZgucryFik?usp=sharing
(future releases should be available from the same link)
Latest version "v16"
-built with latest CUDA 12.2 Update 2 (probably no real speed improvement but why not)
-recompiled everything with gcc 7.3.0. Bernd thinks this might improve validation so i figured I would try it.
To install:
1. Set NNT and deplete your existing cache of tasks (the plan class difference will cause you to nuke/abandon your entire cache if you don't)
2. Stop BOINC
3. Extract the contents of the archive (app_info.xml and and the "BRP7..." executable) to your Einstein project folder.
4. Make sure you preserved the execution permissions of the main "BRP7" executable. if you copy this file with a flashdrive or something, it's likely that permissions will be reset and you will need to set them again with the gui or chmod +x.
5. make sure you check for any missing dependencies on the executable (run it against ldd) and install anything missing
6. Restart BOINC
_________________________________________________________________________
Any chance someone is doing
)
Any chance someone is doing one for the All Sky GW tasks? That would be really nice.
Ben Scott wrote: Any chance
)
not at this time
_________________________________________________________________________
+41
)
+41
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!
V16 runs so far without
)
V16 runs so far without problems on my 4090.
A single wu now takes about 130 seconds, previously ~165 s.
Also the cpu load is lower compared to BRP7 0.16.
Very nice, thanks for that.
Thanks again for sharing the
)
Thanks again for sharing the file. Running on my 1060's, I get a 20% speed boost.
Working great so far. Seeing
)
Working great so far. Seeing >30% improvement on 3080ti.
Ian&Steve C. wrote:To
)
My first try with a petri app so please excuse my ignorance...
I've got a single Ryzen 5700X / RTX 3070 box nearly depleted of E@H GPU WUs.
Can I make the assumption that the "...your entire cache..." refers to cache of GPU E@H WUs only? Not Asteroids/WCG/Prime GPU WUs, not E@H CPU only WUs.
Thanx, Skip
PS: Might have been a bad assumption as far as E@H CPU WUs were concerned.
Wed 25 Oct 2023 03:11:05 PM CDT | Einstein@Home | Didn't resend lost task LATeah1090F_1352.0_6473940_0.0_0 (expired)
yes, deplete your Einstein
)
yes, deplete your Einstein BRP7 cache. the other projects don't matter.
the app_info.xml included will not work for the Einstein CPU tasks and will probably abandon those too.
if you want to continue crunching CPU tasks from Einstein, you'll have to add those parts to the app_info.xml I do not run Einstein CPU tasks, so I couldnt tell you what exactly needs to be added.
_________________________________________________________________________
Wouldn't it simply be a task
)
Wouldn't it simply be a task of saving the E@H cpu apps aside from the BOINC project folder, and then editing the provided app_info.xml to add the cpu apps?
You can get the cpu app details from the client_state.xml file for app_name and plan_class if any and create the cpu sub-section in the app_info file.
Setting the cpu apps aside is to have some safety if the scheduler decides to delete the apps on the first connection with the anonymous platform. If that happens you can copy the cpu apps back into the project folder for the app_info file to use the next connection.
yes that's the general
)
yes that's the general process. i just dont know exactly what needs to be added since I dont run the CPU apps. someone else would have to figure it out
_________________________________________________________________________