1
$\begingroup$

I am confused about what Compile's Listable and Parallelization options do.

(edit: This post is relevant: stack exchange "The option parallelization for compile")

My initial assumption was that Listable was creating multiple threads on a single core. I have 10 cores on my machine. Here is an example.

Parallelization is True:

  cfParallelizationTrue = Compile[{{x, _Real}}, Sin[Sqrt[Abs[x]]],
      CompilationOptions -> {"ExpressionOptimization" -> True}, 
      RuntimeOptions -> {"Speed", "EvaluateSymbolically" -> False},
      RuntimeAttributes -> {Listable}, Parallelization -> True]

With[{rr = RandomReal[{-1, 1}, 1000]},
 RepeatedTiming[cfParallelizationTrue[rr];]]
(*{0.000055932556, Null}*)

Parallelization is False

cfParallelizationFalse = Compile[{{x, _Real}}, Sin[Sqrt[Abs[x]]],
  CompilationOptions -> {"ExpressionOptimization" -> True}, 
  RuntimeOptions -> {"Speed", "EvaluateSymbolically" -> False},
  RuntimeAttributes -> {Listable}, Parallelization -> False]

With[{rr = RandomReal[{-1, 1}, 1000]},
 RepeatedTiming[cfParallelizationFalse[rr];]]
(*{0.000041559204, Null}*)

Parallelization->False gives about 25% speed up. I don't know why.

Does Parallelization mean that it the computation can use multiple kernels?

LaunchKernels[];
$KernelCount
(*10*)
With[{rr = RandomReal[{-1, 1}, 1000]},
 RepeatedTiming[
  ParallelTable[cfParallelizationTrue[rr];, {i, 1, $KernelCount}]]
 ]
(*0.0042045547*)

If there were no overhead for parallel, this should be about the same as running it once on a single core.

With[{rr = RandomReal[{-1, 1}, 1000]},
 RepeatedTiming[
  ParallelTable[cfParallelizationFalse[rr];, {i, 1, $KernelCount}]]
 ]
(*0.0041659531*)

There seems to be no benefit in any case for using multiple kernels.

Is this because each call to cfParallelizationFalse[rr] is using all my cores because cfParallelizationFalse has Listable->True even though there is only one kernel??

And, cfParallelizationTrue[rr] in a ParallelTable is using all the kernels, but they are running 10 times slower because each Kernel's core is busy ??

Improve this question
$\endgroup$
5
  • 5
    $\begingroup$ Listable means the function is threaded over lists. Compiled functions are different from Regular Listsble functions in that the threading depends on the argument type. A compiled function with a vector argument will thread down to vectors; a regular functions always threads down to scalars $\endgroup$
    – Michael E2
    Aug 6, 2022 at 17:10
  • $\begingroup$ Parallelization in the WVM which requires Listable has minimal overhead or at least optimized overhead, I believe. Due to restricted data types and data sharing. $\endgroup$
    – Michael E2
    Aug 6, 2022 at 17:20
  • 1
    $\begingroup$ Parallelizing a parallelized process sounds like compressing a compressed file— the result is almost always worse. But I don’t know that about parallelization $\endgroup$
    – Michael E2
    Aug 6, 2022 at 17:24
  • $\begingroup$ I think part of my confusion stems from not understanding what a multi-threaded process on a single core does, as opposed to multi-threaded processes on multiple cores. I was under the assumption that each launched kernel was attached to one of the cores—and each of those cores could be multi-threaded. Perhaps Threadable in the WL has nothing to do with running a multi-thread process. $\endgroup$
    – Craig Carter
    Aug 6, 2022 at 21:14
  • 4
    $\begingroup$ (Thread[]](reference.wolfram.com/language/ref/Thread.html) and process thread are unrelated concepts. Listable is related to the first and not the second. -- Parallelization in the WVM is handled without launching separate kernels, unlike ParallelTable[] and other Parallel* functions. I don't believe kernels are attached to specific cores, but there's usually little point in having more kernels than cores. I think the scheduling of processes is left to the OS, but that's my assumption. $\endgroup$
    – Michael E2
    Aug 6, 2022 at 21:57

0

Reset to default

Your Answer

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

Browse other questions tagged or ask your own question.