Particle creation is slow in parallel simulatins

[RudolfWeeber]

The time for creating a particle scales linearly in the number of mpi ranks for high core numbers (>~64)
E.g 30 ms per particle creatoin (explicitly specifying the particle id) for 512 MPi ranks on the ant cluster at ICP.

In my opionion, particle insertoin should comprise

• a broadcast from Python on rank 0 to all ranks (scales n log n in number of ranks)
• maybe some point to point communicaoitn at constatn time

Steps to investigate

Use a Python profiler on the particle creation code in src/python/espressomd.particle_data.py with different number of mpi ranks to figure out, where in the creation process the poor scaling appears.

The script was

import espressomd
import numpy as np
from time import time

s=espressomd.System(box_l=[1,1,1])
grid =s.cell_system.node_grid
n=10*np.prod(grid)
print(grid)
for n in 1,10,100:
    tick=time()
    s.part.add(id=range(n),pos=np.random.random((n,3))*s.box_l)
    tock=time()
    print(n, (tock-tick)/n)
    s.part.clear()

And the commands to run were

module load spack/default gcc/12.3.0 cuda/12.3.0 openmpi/4.1.6 \
            fftw/3.3.10 boost/1.83.0 cmake/3.27.9 python/3.12.1

`
srun --cpu-bind=cores --ntasks=512 --time=60 -u ./pypresso ./test.py