Are CPU Work Units also crunched on GPUs?
9 Jul 2024 4:57:44 UTC
Topic 231283
(moderation:
A simple question with perhaps a complex answer. We know GPUs are much more efficient with most tasks, which has me wondering, is CPU work for this project (and Asteroids@home) superfluous because the work done can also be created as GPU work units (like on good 'ole SETI@home)? Or is CPU work on this platform necessary to be done on CPU? Do CPU WUs have a necessary role or are they space-heaters adjacent?
My farm is all CPUs so i attach to projects that need CPU muscle, otherwise should be done on efficient silicon (GPUs).
Folding is clear which sub projects are designated for which silicon and the quantities which makes it clear they have a role there, and CPU-only BOINC projects of course.
Thank You!
Edit; Ah.
https://einsteinathome.org/apps.php
FGRP5 BRP7 BRP4G O3AS BRP4 BRP4A FGRPB1G
CPU GPU CPU GPU CPU-ish ARM GPU
Note* GPU WUs do need some CPU compute, but GPU is necessary aspect. It does look like I have answered my own question. CPU-only WUs FGRP5 & BRP4G!
Computers installed; https://mega.nz/folder/RwVDSYRL#Lj6-eHQLTMvX2nMLuiGTxw
If you visit your einstein
)
If you visit your einstein project preferences, it will state which applications can run on cpu or gpu or both. Well some of the applications already ended but still show up in the project preferences list. For e.g. the "Multi-Directional Gravitational Wave search on O3" has both cpu and gpu tasks. Both ended and continue with the gpu only application "All-Sky Gravitational Wave search on O3". For these three applications that I mentioned, all I know is that we are processing the data collected on the third observation (called O3, or Observation no. 3) of the ground base detectors. I don't know the difference if any between the "Multi-Directional" versus the "All-Sky".
Other than gravitational wave search, there are radio wave (low frequency) and gamma-ray (high frequency) pulsar search of various ground base detectors such as form the Arecibo observatory (no longer operational) and Meerkat radio telescope. Data collected can be processed in either cpu/gpu or gpu only. Obviously gpu will be much faster.
APPLICATIONS
Run only the selected applications
Binary Radio Pulsar Search (Arecibo, GBT, CPU/GPU) [BRP4]
Binary Radio Pulsar Search (Arecibo, GBT, long, CPU) [BRP4G]
Binary Radio Pulsar Search (Arecibo, GBT, arm64, CPU/GPU) [BRP4A]
Binary Radio Pulsar Search (MeerKAT, GPU) [BRP7]
Gamma-ray pulsar search #5 (CPU) [FGRP5]
Gamma-ray pulsar binary search #1 (GPU) [FGRPB1G]
All-Sky Gravitational Wave search on O3 (GPU) [O3AS]
Multi-Directional Gravitational Wave search on O3 (GPU) [O3MDF]
Multi-Directional Gravitational Wave search on O3 (CPU) [O3MD1]
Thank you! So the question
)
Thank you! So the question is, can
Binary Radio Pulsar Search (Arecibo, GBT, long, CPU) [BRP4G] and
Gamma-ray pulsar search #5 (CPU) [FGRP5] work be compiled instead for a primarily GPU workload? It looks like that is not the case or not being done, x86 CPU core crunchers are necessary for these 2 so jives with my farm mission.
Computers installed; https://mega.nz/folder/RwVDSYRL#Lj6-eHQLTMvX2nMLuiGTxw
Probably yes and getting
)
Probably yes and getting programming resources to port the cpu code to gpu code. In a nutshell, we are doing signal processing analysis (FFT) type of work. Signal processing has high degree of parallelism in computation, so my opinion is that it could be possible to move signal processing work to gpu. I don't how know complex it is to do this here for FGRP5.
For example about gramma-ray pulsar search, from what I read here , there are apps for both cpu and gpu (slide #3) when it comes to gamma-rays pulsar search. Specifically when searching for binary pulsars (slide #14), the original application was in cpu (FGRPB1) and then later ported to gpu (FGRPB1G), "G" for gpu, I assume.
There is a lot to read about Einstein@Home and there are many publications if you are interested. From what I know this project is very well supported by the Einstein@Home team.
Edit: both FGRPB1 and FGRPB1G already ended.
Wonder if someone can give a
)
Wonder if someone can give a high-level reason why BRP4G and FGRP5 are not coded for GPU-centric crunching. Is there a technical reason, or to be inclusive for non-gaming PCs? Thanks
Computers installed; https://mega.nz/folder/RwVDSYRL#Lj6-eHQLTMvX2nMLuiGTxw
BRP4 and BRP4G are the same
)
BRP4 and BRP4G are the same tasks.
Just that the G tasks for gpus are 8 bundled BRP4 tasks instead of just one.
Just sent to different devices IF the application is configured for them.
I run the same app for both types of task with a custom app and the anonymous app platform.
https://einsteinathome.org/host/12965439/tasks/0/0
https://einsteinathome.org/host/12965439/tasks/0/19
For some reason they named it
)
For some reason they named it such that BRP4 is GPU, BRP4G is CPU, according to the https://einsteinathome.org/apps.php. I see my basic (2005 pentium m) PC is running BRP4G 1.33 on it now.
So work units of the same format and algorithm can be divided up for GPU or CPU depending on loads i guess.
Server status page shows 158 days until BRP4 completion, not long!
Looks like FGRP5 with 2377 days remaining is the 1 CPU-only subproject long term?
Computers installed; https://mega.nz/folder/RwVDSYRL#Lj6-eHQLTMvX2nMLuiGTxw
there is a lot of confusion
)
there is a lot of confusion in the BRP4 vs BRP4G tasks because things have changed a few times over the years. so the way things are running now is not the way it was originally intended.
initially the 'G' tasks were intended for GPUs, and as Keith said, they were bundled. same work, just more of it. i don't remember what exactly the reason, but as some point they stopped running BRP4G and threw everything back to the BRP4 pool and everyone was working on the same stuff.
then a few years ago, due to both heavy load on the BRP4 server and poor validation from Arm and Intel-GPU devices, they split it back up into BRP4 and BRP4G, and put Intel-GPU devices on BRP4G.
then they did a psuedo-rename, where BRP4G became "Arecibo,GBT,long", and BRP4 was "Arecibo,GBT". you can see these labels in the applications page. though the BRP4G and BRP4 labels still exist within the BOINC client/server to delineate them. (too difficult and troublesome to change in the BOINC infrastructure)
and then again they moved all intel-GPUs and Arm devices to the "short/BRP4" and left normal CPUs to the "long/BRP4G" project, again i think due to validation issues. but it's hard to keep track. the BRP4 projects have always had various validation issues across all the device types in same way it seems.
yes, gamma ray #5 is CPU only. keep in mind that all of the subprojects are technically crunching different stuff.
_________________________________________________________________________