2
$\begingroup$

I have programs which can not use CUDA (graphic cards), but can run perfectly in parallel using ParallelTable (mainly symbolic transformations). Now I wonder, whether ParallelTable works with co-processors, such as the Intel Xeon Phi with 61 cores. If I would have such a co-processor, could I start 61 additional instances of Mathematica which then run my programs on the co-processors?

I was looking here, here and here, but was not able to find the information.

$\endgroup$
  • $\begingroup$ Unfortunately I do not have an answer regarding the usage of co-processors. But I want to advice you (if you haven't done so already) to test parallelizability of your program with a multi-core processor and check for speed improvements over single core execution speeds. In practice parallel execution has to deal with data transfers, that are often neglected when theorizing about parallel execution. I had this multiple times when I thought parallel execution should yield speed improvements but the overhead led to not having any benefits from it at all. $\endgroup$ – Wizard Nov 24 '16 at 13:57
  • $\begingroup$ @Wizard thank you for the hint. Yes I tested it with my 8-core maschine, and got a speedup of roughly 8 compared to the single-maschine version. The code uses very little data, but a lot of computation that's why i think it worked. $\endgroup$ – NicoDean Nov 24 '16 at 14:01
  • 2
    $\begingroup$ No, there is no support for the Xeon Phi or other coprocessors. $\endgroup$ – ilian Nov 25 '16 at 16:29
  • $\begingroup$ @ilian thanks for your answer. I found an official spec-sheet from HP which mentions that Wolfram Mathematica should be supported (www8.hp.com/h20195/v2/getpdf.aspx/c04168329.pdf) could you please tell me what you think about this. Thanks a lot! $\endgroup$ – NicoDean Nov 26 '16 at 21:53
  • 2
    $\begingroup$ @NicoDean It is, unfortunately, inaccurate. The kernel has not been ported to the mic architecture, so it will not run natively on those cores -- hence ParallelTable and friends cannot work. Automatic offloading using MKL is somewhat more feasible but the results from in-house testing have been underwhelming. $\endgroup$ – ilian Nov 27 '16 at 16:10
4
$\begingroup$

From the comments,

No, there is no support for the Xeon Phi or other coprocessors.

The kernel has not been ported to the mic architecture, so it will not run natively on the Xeon Phi cores -- hence ParallelTable and friends cannot work.

Automatic offloading for certain linear algebra operations using Intel's MKL library is somewhat more feasible but the results from in-house testing have been underwhelming.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.