Performance Optimisation and Productivity
A Centre of Excellence in HPC
Communication imbalance initial analysis
The amount of communication imbalance exposed by the Communication Imbalance kernel will depend on how the neighbor lists are initialized.
The following Figure shows an example of a 24 processors run in Marenostrum IV where every process communicates to approximately 8 others plus to rank 0, according with the original version of the algorithm.
The trace shows how the imbalance in the number of MPI_Irecvs and MPI_Sends
done by rank 0 propagates through the balanced compute and is paid by all other
threads in the MPI_Allreduce.
