Bug in refreshIDs using mpt on HAWK about ls1-mardyn HOT 3 CLOSED

Can you provide some more information:

• how you configured and build ls1 using Makefile or cmake?
• which mpt and compiler version did you use?
• what was the exact mpirun command to start ls1?

from ls1-mardyn.

Sure:
I used the default configuration but changed the vectorization to AVX2 in the Makefile
VECTORIZE_CODE ?= AVX2
Later I tried out the debug version, which didnt work either.
Compiled with
make clean && make -j -f Makefile in the src folder after loading the default modules module load gcc mpt

This results in the modules gcc/9.2.0 and mpt/2.23 to be loaded.

The compilation info of ls1 prints:

INFO:   20201202T124202 0.000435 [0]    Compilation info:
INFO:   20201202T124202 0.000455 [0]            Compiler:       GNU 9.2.0
INFO:   20201202T124202 0.000471 [0]            Compiled on:    Dec  2 2020 08:16:51
INFO:   20201202T124202 0.000485 [0]            Precision:      Double (DPDP)
INFO:   20201202T124202   0.0005 [0]            Intrinsics:     AVX2
INFO:   20201202T124202 0.000514 [0]            RMM/normal:     normal mode. All features work in this mode.
INFO:   20201202T124202 0.000535 [0]            OpenMP:         without OpenMP support.
INFO:   20201202T124202 0.000551 [0]            MPI:            with unknown MPI (version 3.1)

Submitted the following script to the test queue qsub -q test hawk_start.sh :

#! /bin/bash
#PBS -N mpt2
#PBS -l select=2:node_type=rome:mpiprocs=128
#PBS -l walltime=00:10:00
module load gcc mpt
cd $PBS_O_WORKDIR
mpirun -np 256 ../../ls1-mardyn/src/MarDyn.PAR_DEBUG_AVX2_mpt config.xml --final-checkpoint=0

from ls1-mardyn.

@cniethammer that function should probably better be implemented using MPI_Exscan, what do you think?
Currently, it is a sequential communication from rank 0 to rank 1 to rank 2, ... to rank N-1.

But it should still work...

from ls1-mardyn.