Parallelization in MAPDL

[akaszynski]

Posting an email received:

I am making a simple example to see if we can parallelise due to python threading/multiproc. capabilities. My script is primitive, and not sure if it works well in parallel mode (with out the parallel processes the main () takes about 15 sec. Was wondering if you have an example where we parallelize and execute things (methods) at the same time (assuming it is possible with mapdl)?

import multiprocessing
from ansys.mapdl.core import launch_mapdl

mapdl = launch_mapdl()

def klines():
    mapdl.run('/PREP7')
    mapdl.run('ADELE,ALL')
    mapdl.run('LDELE,ALL')
    mapdl.run('KDELE,ALL')

    for i in range(1,500,1):
        mapdl.parameters['Kx']=i
        #mapdl.run('K, Kx, 0, 0, 0')
        mapdl.k('', i, 0, 0)

    #mapdl.kplot()

def klinestwo():

    mapdl.run('/PREP7')
    mapdl.run('ADELE,ALL')
    mapdl.run('LDELE,ALL')
    mapdl.run('KDELE,ALL')

    for i in range(501,1000,1):
        mapdl.parameters['Kx']=i
        #mapdl.run('K, Kx, 0, 0, 0')
        mapdl.k('', i, 0, 0)

    #mapdl.kplot()

def main():

    t1=time.time()

    process1 = multiprocessing.Process(target=klines)
    process2 = multiprocessing.Process(target=klinestwo)
    process1.start()
    process2.start()
    process1.join()
    process2.join()
    #klines()

    #klinestwo()
    t2=time.time()
    print('finished:  '+str(t2-t1))

main()

[akaszynski]

Solution isn't to use multiprocessing. MAPDL is limited by the fact that the head process is single threaded, so any preprocessing commands will be run sequentially. Running two processes in parallel will potentially result in garbled output. Same with post-processing.

The trick here is to use mapdl.non_interactive where you can effectively send all the looped commands in one batch to mapdl in one go, but still use Python loops:

Example

from ansys.mapdl.core import launch_mapdl

mapdl = launch_mapdl()

def klines():
    mapdl.clear()
    mapdl.prep7()

    for i in range(1, 500):
        mapdl.k('', i, 0, 0)


def klines_non_interactive():
    with mapdl.non_interactive:
        mapdl.clear()
        mapdl.prep7()
        for i in range(1, 501):
            mapdl.k('', i, 0, 0)

Timings

>>> timeit klines()
207 ms ± 4.05 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> timeit klines_non_interactive()
52.9 ms ± 81.2 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Additional thoughts

I'd really prefer to have a lower level interface to the keypoints so we could just pass a large batch of them as an array. I doubt that's a feature we're going to focus on, but if enough users are still using PREP7 commands within MAPDL, we can shift focus to that.