nimforum mirror - Parallel processing helper procs?

jlp765 (orginal) [2017-12-12T01:43:28+01:00] view original

Should Nim make it easier to perform the basic use-cases of parallelism (i.e., abstract away the looping)?

Three main use-cases that I am thinking of:

Run something in parallel, that may or may not use the loop number

Run something in parallel that returns a result (and probably does use the loop number)

Accumulate the result of use-case 2. (i.e., doesn't need to return each value, just some resultant single value)

The third use-case would not be implemented, but can be performed by applying existing sequence accumulators to use-case 2.

My thoughts were around functional-programming style helper things like the following (using for example, 42 as the number of times to run these in parallel, i.e. 0..41 ):

runParallel(42, myProcToRunParallel(....))

var res: seq[int] = runParallel[int](42, someResultReturningProc(.....))

If the "it" style syntax was used, then example 2. would be

var res: seq[int] = runParallelIt[int](42, someResultReturningProc(.., 5*it, ....))

where "it" would be 0..41

I don't understand the complexities of parallelism, but would this be possible, and would it be a good thing to implement if it is possible?

cdome (orginal) [2017-12-12T13:26:51+01:00] view original

Have seen Nim's parallel statement? IMO, it does exactly what you have described. https://nim-lang.org/docs/manual.html#parallel-spawn-parallel-statement

mratsim (orginal) [2017-12-12T14:38:09+01:00] view original

There are several types of parallelism.

Quoting myself

data parallelism: executing the same instructions on multiple cores

task parallelism: multiple different tasks/functions on multiple cores

model parallelism: this is somewhat specific to science, if you need to evaluate A and B and they are independant, evaluate both in parallel.

In case of data parallelism, if it's done on a shared memory data structure there are non-trivial issues to solve that are better done through OpenMP or via a library like Rust rayon. Typical issues are static or dynamic scheduling, false sharing/CPU cache contention, race conditions.

In case of task parallelism, you can use parallel and spawn mentionned by cdome or channels.

Mirror of forum.nim-lang.org

3413 :: Parallel processing helper procs?