cselab / korali_benchmark Goto Github PK

The usage of pure MPI on a cluster of multi-core nodes is a very bad approach.
If it is possible, a hybrid MPI+Threads approach must be used, i.e., one process with multiple worker threads per node.
Otherwise, a server thread (or asynchronous progress thread) is used in each process, which leads to core oversubscription.
This is a well-known issue of pure MPI code, which is a naive and outdated approach.
Therefore, the experimental benchmarking of pi4u/torc framework is wrong.