PoCL 5.0-RC1 Released With Experimental OpenCL For Networked Systems

[phoronix]

Phoronix: PoCL 5.0-RC1 Released With Experimental OpenCL For Networked Systems

PoCL 5.0-RC1 is out today as the newest feature release being brewed for this "Portable Computing Language" implementation that allows for OpenCL code to run on CPUs as well as running OpenCL code on other back-ends such as atop NVIDIA CUDA and AMD ROCm and other LLVM back-ends...

https://www.phoronix.com/news/PoCL-5.0-RC1

[Mathias]

PoCL has ROCm and CUDA backends? This is getting interesting: PoCL vs RustiCL vs some AMD OpenCL implementation (aren't there more then one?)

[illwieckz]

That's very good! I don't get why this is not a thing since already a decade! I tried to revive CLara but the task of implementing and maintaining such “OpenCL over the network” feature will be better done by PoCL guys.

So, soon we will be able to select in OpenCL apps GPUs and CPUs from other computers in the network. Actually it will give usefulness to CPU implementations of OpenCL as an easy way to dispatch compute on other computers. In the mean time, Blender still doesn't allow to mix GPUs from different hardware brand to compute a single render, because you have to choose if you want to do CUDA, OneApi or HIP, and only one of them at once.

[kpedersen]

That's very good! I don't get why this is not a thing since already a decade!

There kind of has been but it keeps getting broken. People's need for it tends to come in waves and it gets neglected again and starts to rot until the next time.

My thesis was basically to implement a distributed network aware OpenGL (kind of the opposite to VirtualGL). I was amazed at how so much of this stuff existed but in such a derelict form.

[illwieckz]

There kind of has been but it keeps getting broken. People's need for it tends to come in waves and it gets neglected again and starts to rot until the next time.

That's sad, but yeah…

My thesis was basically to implement a distributed network aware OpenGL (kind of the opposite to VirtualGL). I was amazed at how so much of this stuff existed but in such a derelict form.

Very interesting, do you have a link to that thesis document?

[Eirikr1848]

Yet again I am excited but confused.

Does this allow the simultaneous use of multiple OpenCL backends for networked parallel processing?

Utilization of all available resources in a network, including a shared VRAM pool for tasks that exceed a single or dual local GPU setup is particularly fascinating.

Very curious how the scheduler works to determine how to best distribute the workload... Especially considering factors like the computational capabilities / CL extensions / FP Performance / etc of each node and network latency.

[coder]

Utilization of all available resources in a network, including a shared VRAM pool for tasks that exceed a single or dual local GPU setup is particularly fascinating.

Not sure exactly what you're asking, but it's certainly not going to support SVM across the network.

Very curious how the scheduler works to determine how to best distribute the workload... Especially considering factors like the computational capabilities / CL extensions / FP Performance / etc of each node and network latency.

Last I checked, OpenCL didn't do load-balancing for you. That's still left up to the application.

[Eirikr1848]

Not sure exactly what you're asking, but it's certainly not going to support SVM across the network.


Last I checked, OpenCL didn't do load-balancing for you. That's still left up to the application.

That's what I was wondering, thank you. It made no sense to have network SVM and auto-load balancing. Which would be cool, albeit impractical.

Thanks for clarifying

[coder]

That's what I was wondering, thank you. It made no sense to have network SVM and auto-load balancing. Which would be cool, albeit impractical.

First, take what i say with a grain of salt. Pekka is one of the authors and would be able to answer your questions with certainty.

Second, it's probably not impossible to have an OpenCL stack do some sort of load-balancing. I'm guessing you could make a special kind of device that maps to a collection of devices or network nodes. However, I'd expect most developers would be able to load-balance and pipeline more effectively - especially since data-movement could be a bottleneck if the load balancer is too naive.

Finally, it's probably most helpful to consider what this stands to replace, which is probably low-level APIs like MPI (Message-Passing Interface). Compared to that, OpenCL i a big step up!

[kpedersen]

Very interesting, do you have a link to that thesis document?

Sure, you can grab it here:

https://eprints.bournemouth.ac.uk/36043/1/PEDERSEN%2C%20Karsten_Ph.D._2021.pdf

That is quite long winded. You might also prefer the more succinct paper here on just the networking part:

https://eprints.bournemouth.ac.uk/31543/1/aamua.pdf

Some ideas were good, some were less so. Let me know if you have any questions