Print

[unaimatakenny]

Hello,

I have one year using JPPF, and I wonder if JPPF team plan to use the processing power of GPUs for parallel processing of tasks, like JCUDA.

I would like to know the pros and cons you can have in developing this type of feature? If compatible with the architecture of JPPF? ...

I think it's an interesting topic to consider, since the GPUs have more cores than current CPUs.

Thank you very much,

[lolo]

Hello,

We have not yet considered using GPU programming.
Unfortunately, I'm not familiar enough with this paradigm to even tell if it is feasible with JPPF.
Please, feel free to enter a feature request for this, giving any detail you can think of, this would be very much appreciated and would give us a way to track this discussion over time.

Sincerely,
-Laurent

[rickl]

Hello,

Sorry I know this thread is old but I am also interested in the possibility of GPU/JCUDA integration with JPPF.  I was unable to vote for it on the issue tracker (no account) so please consider this a vote for the Feature Request JPPF-12.

Thank you for all of you efforts on the JPPF project!

Sincerely,
-Rick

[lolo]

Hi Rick,

Thank you very much for your support of this feature.
We have planned to do something for the next release (v3.3). Given the state of what we know about GPU  programming, this will most likely be experimental and probably delivered as a sample.
If you would like an account for the issue tracker, please shoot me an email at laurent.cohen at jppf.org and I will set one up for you. I understand this is not very convenient, but this is the only way I found to avoid spambots messing with us 

Sincerely,
-Laurent

[Alexx]

Hello,

Please add my vote to the Feature request JPPF-12. I write here as I don't have an account at the Issue Tracker.
The Rootbeer project you mentioned at the issue tracker looks very promising.
I found 2 interesting papers on Rootbeer here https://github.com/pcpratts/rootbeer1/tree/master/doc,
one of them has speed comparisons.

It would be a great if JPPF could utilize GPUs as well, so users save huge amounts of time not having to learn the particulars of GPU programming themselves.

I did some searching and found this:
1) JOCL (Java bindings for OpenCL) http://www.jocl.org/
2) APARAPI http://developer.amd.com/blog/i-dont-always-write-gpu-code-in-java-but-when-i-do-i-like-to-use-aparapi/
http://code.google.com/p/aparapi/
the latter page contains links to the following articles (among other potentially useful references):

Similar Work
Peter Calvert's java-GPU has similar goals and offers a mechanism for converting Java code for use on the GPU
Check out Peter's dissertation "Parallelisation of Java for Graphics Processors" which can be found here
Marco Hutter's Java bindings for CUDA
Marco Hutter's Java bindings for OpenCL
Ian Wetherbee's Java acceleration project - creates accelerated code from Java (currently C code and native Android - but CUDA creation planned)
"Rootbeer: Seamlessly using GPUs from Java" by Philip C. Pratt-Szeliga

And lastly, here's a video from AMD's official youtube channel:
"Performance Evaluation of AMD-APARAPI Using Real World Applications"
http://www.youtube.com/watch?v=o0ZYz9lrp_Y
and the official AMD APARAPI page http://developer.amd.com/tools/heterogeneous-computing/aparapi/

So, it looks to me that there are at least two other alternatives: JOCL and APARAPI.

I hope this helps.

Alex.

PS
Could you shed some light as to where the development of JPPF stands at this point in terms of GPU utilization?
Thank you in advance.

[lolo]

Hello Alex,

Sorry for this late response.
Thank you very much for the links and pointers, this is very helpful.
After reviewing these, I found that APARAPI is very attractive to a Java developer like me. I especially like the fact that the APARAPI code will execute even if the OpenCL bindings are not present, albeit slower of course.

So where we are currently in terms of GPU utilization, is not very far. However, our plan is to introduce a GPU computing sample in the upcoming JPPF 3.3 (planned for end of March - 1st half of April).
Given the programming model for GPU computing, I don't think there will be a need for specifc APIs in JPPF. It should be a simple matter of using APARAPI from within a JPPF task.

Any input you have will be much appreciated.

Sincerely,
-Laurent