MapdlPool slower than launch_mapdl

[germa89]

Gather information about a reported issue with MapdlPool speed.

Notes

All the results use v222 on Windows.

Pool uses 8 cores, which are the physical cores of my laptop.

Results

Iteration	launch_mapdl	MapdlPool
1	141.575	487.243
2	149.290	487.936
3	156.204	488.373
4	153.430	489.259
5	153.133	489.637
6	155.844	490.195
7	154.454	491.174
8	155.081	493.576
9	152.449	469.160
10	149.174	472.584
11	153.367	472.701
12	149.009	481.007
13	158.336	481.815
14	159.143	483.254
15	159.136	480.907
16	158.764	479.020
17	161.264	162.742
18	160.214	164.233
19	159.055	163.783
20	159.422	157.527
Total	3117.357	1137.972

MapdlPool result

Total time: 1137.9721553325653
[487.2430682182312, 487.93626523017883, 488.3732578754425, 489.2591965198517, 489.63740062713623, 490.1950089931488, 491.1746802330017, 493.5765743255615, 469.1609597206116, 472.5841302871704, 472.7011818885803, 481.007390499115, 481.8152985572815, 483.2545311450958, 480.9073061943054, 479.0204613208771, 162.74245071411133, 164.2335982322693, 163.7831094264984, 157.52774477005005]

Full console output

Pool: True
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...

PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
Total Time 487.2430682182312
Total Time 487.93626523017883
Total Time 488.3732578754425
Total Time 489.2591965198517
Total Time 489.63740062713623
Total Time 490.1950089931488
Total Time 491.1746802330017
Total Time 493.5765743255615
Total Time 469.1609597206116
Total Time 472.5841302871704
Total Time 472.7011818885803
Total Time 481.007390499115
Total Time 481.8152985572815
Total Time 483.2545311450958
Total Time 480.9073061943054
Total Time 479.0204613208771
Total Time 164.2335982322693
Total Time 162.74245071411133
Total Time 163.7831094264984
Total Time 157.52774477005005

Total time: 1137.9721553325653
[487.2430682182312, 487.93626523017883, 488.3732578754425, 489.2591965198517, 489.63740062713623, 490.1950089931488, 491.1746802330017, 493.5765743255615, 469.1609597206116, 472.5841302871704, 472.7011818885803, 481.007390499115, 481.8152985572815, 483.2545311450958, 480.9073061943054, 479.0204613208771, 162.74245071411133, 164.2335982322693, 163.7831094264984, 157.52774477005005]

launch_mapdl result

Total time: 3117.3571031093597
[141.57540607452393, 149.29076719284058, 156.20481204986572, 153.43074774742126, 153.1332221031189, 155.84478402137756, 154.45419192314148, 155.0814709663391, 152.44991850852966, 149.1741588115692, 153.3673119544983, 149.00977158546448, 158.3366515636444, 159.14314198493958, 159.1366572380066, 158.76436352729797, 161.26456499099731, 160.21481370925903, 159.0556252002716, 159.42233037948608]

Full console output

Pool: False
PyMAPDL is taking longer than expected to connect to an MAPDL session.
Checking if there are any available licenses...
Solving 0 model.
Total Time 141.57540607452393
Solving 1 model.
Total Time 149.29076719284058
Solving 2 model.
Total Time 156.20481204986572
Solving 3 model.
Total Time 153.43074774742126
Solving 4 model.
Total Time 153.1332221031189
Solving 5 model.
Total Time 155.84478402137756
Solving 6 model.
Total Time 154.45419192314148
Solving 7 model.
Total Time 155.0814709663391
Solving 8 model.
Total Time 152.44991850852966
Solving 9 model.
Total Time 149.1741588115692
Solving 10 model.
Total Time 153.3673119544983
Solving 11 model.
Total Time 149.00977158546448
Solving 12 model.
Total Time 158.3366515636444
Solving 13 model.
Total Time 159.14314198493958
Solving 14 model.
Total Time 159.1366572380066
Solving 15 model.
Total Time 158.76436352729797
Solving 16 model.
Total Time 161.26456499099731
Solving 17 model.
Total Time 160.21481370925903
Solving 18 model.
Total Time 159.0556252002716
Solving 19 model.
Total Time 159.42233037948608
Total time: 3117.3571031093597
[141.57540607452393, 149.29076719284058, 156.20481204986572, 153.43074774742126, 153.1332221031189, 155.84478402137756, 154.45419192314148, 155.0814709663391, 152.44991850852966, 149.1741588115692, 153.3673119544983, 149.00977158546448, 158.3366515636444, 159.14314198493958, 159.1366572380066, 158.76436352729797, 161.26456499099731, 160.21481370925903, 159.0556252002716, 159.42233037948608]

Code

Do change the variable POOL value to True or False to switch between MapdlPool and launch_mapdl.

Details

import time

import numpy as np

POOL = False
solve = True

exec_file = r"C:\Program Files\ANSYS Inc\v212\ansys\bin\winx64\ANSYS212.exe"
exec_file = r"C:\Program Files\ANSYS Inc\v222\ansys\bin\winx64\ANSYS222.exe"
# exec_file = r"C:\Program Files\ANSYS Inc\v231\ansys\bin\winx64\ANSYS231.exe"


esize = 10.
length = 100.
width = 100.
depth = 100.
numsteps = 1

global total_time
total_time = []

def solve_model(mapdl, arg=None):
    mapdl.clear()
    if solve:
        mapdl.fcomp('rst', 0)
        # Write all data in double precision
        mapdl.run('/config,resuprec,0')
        mapdl.prep7()
        mapdl.k(1, 0.0, 0.0, 0.0)
        mapdl.k(2, length, 0.0, 0.0)
        mapdl.k(3, length, width, 0.0)
        mapdl.k(4, 0.0, width, 0.0)
        mapdl.k(5, 0.0, 0.0, depth)
        mapdl.k(6, length, 0.0, depth)
        mapdl.k(7, length, width, depth)
        mapdl.k(8, 0.0, width, depth)
        mapdl.v(1, 2, 3, 4, 5, 6, 7, 8)

        mapdl.et(1, 186)
        mapdl.mp('ex', 1, 210000.)
        mapdl.mp('prxy', 1, .3)
        mapdl.mp('kxx', 1e-6)

        mapdl.esize(esize)
        mapdl.mshape(0, "3D")
        mapdl.vmesh('all')
        mapdl.finish()

        mapdl.slashsolu()
        # For some reason - with d,all,all - "ENS" block doesn't appear in XPL
        mapdl.d('all', 'ux', 2.0)
        mapdl.d('all', 'uy', 2.0)
        mapdl.d('all', 'uz', 2.0)
        mapdl.bf('all', 'temp', 100.)
        mapdl.time(1.)
        mapdl.nsubst(numsteps, numsteps, numsteps)
        mapdl.outres('all', 'none')
        mapdl.outres('strs', 'all')
        mapdl.outres("epel", "all")
        mapdl.outres("eppl", "all")
        mapdl.outres("nsol", "all")
        mapdl.solve()
        mapdl.finish()

        mapdl.aux2()
        mapdl.combine('FULL')
        mapdl.finish()

    xpl = mapdl.xpl
    mm = mapdl.math

    t0 = time.time()
    xpl.open('dummy_solve.rst')

    header = xpl.read('HEADER', asarray=True)
    # print(header)
    num_nodes = header[1]
    num_elems = header[5]
    xpl.close()

    # Get internal Element order
    xpl.open('dummy_solve.rst')
    internal_elem_order = xpl.read("ELM", "ELEM_ORDER")
    xpl.close()

    stress = [float(id) for id in range(1, 2001)]

    all_time_steps = []

    def get_elem_stress(eid):
        return stress[eid - 1]

    t0 = time.time()
    for i in range(1, numsteps + 1):
        t_elem = 0.
        for u_eid in range(1, num_elems + 1):
        # for index, eid in enumerate(internal_elem_order):
            i_eid = internal_elem_order[u_eid - 1]
            xpl.open(r"dummy_solve.rst", "WRIT")
            xpl.step(f"DSI::SET{i}::ESL::ELEM{u_eid}")
            t0_elem = time.time()
            stress_elem = get_elem_stress(i_eid)
            myvec = np.ones(48, dtype=np.float64) * float(i * stress_elem)
            mm.set_vec(myvec, 'NEWVEC')
            xpl.write('ENS', 'NEWVEC')
            t1_elem = time.time()
            t_elem = t1_elem - t0_elem
            all_time_steps.append(t_elem)
            xpl.up()
            xpl.close()

    t1 = time.time()

    print(f"Total Time {t1 - t0}")
    total_time.append(t1-t0)


print(f"Pool: {POOL}")
if POOL:

    from ansys.mapdl.core import MapdlPool
    pool = MapdlPool(
        8,
        start_instance=True,
        exec_file=exec_file,
        jobname='dummy_solve',
        override=True,
        start_timeout=90,
        loglevel="DEBUG"
        )

    full_0 = time.time()
    pool.map(func=solve_model, iterable=[None for each in range(20)])
    full_1 = time.time()
    pool.exit()

else:
    from ansys.mapdl.core import launch_mapdl
    
    model_path = r"C:\Users\gayuso\pymapdl\tmp\Ayush\pymapdl"
    mapdl = launch_mapdl(exec_file, run_location=model_path, jobname='dummy_solve', override=True)

    full_0 = time.time()

    for i in range(20):
        print(f"Solving {i} model.")
        solve_model(mapdl, None)
    
    full_1 = time.time()
    mapdl.exit()

print(f"Total time: {full_1-full_0}")
print(total_time)

[germa89]

It is strange that the last four iterations take less time than the previous 16 (8x2). It seems that PyMAPDL is not parallelizing correctly.

In the case of launch_mapdl the time spent is comparable (150s).

[germa89]

It seems it is not always performant to use all the CPUs when using multi-threading:

Reference: https://www.dabeaz.com/python/UnderstandingGIL.pdf