METIS error message
Not sure it helps but might give some hints. I am running the small cosmological volume (2M particles) on 1024 nodes of JUQUEEN. On the machine, there is no installation of METIS available, only parMETIS. I use the version 4.0.2.
When running with the following options
runjob --ranks-per-node 1 : ./test_fixdt_mpi -r 100 -t 32 -f CosmoVolume/cosmoVolume.hdf5 -m 0.2 -w 5000 -d 1e-8
I get the following "error" messages when doing the domain decomposition:
***Cannot bisect a graph with 0 vertices! ***You are trying to partition a graph into too many parts!
which apparently comes from parMETIS itself. However, the code keeps running even with some nodes with 0 particles and 0 cells. Could it be that parMETIS has corrected the bug we suspect is present in METIS ?