just having a look at some MPI collective communication results; in particular, how quick different implementations of some common MPI collectives are in relation to each other.
I explore 7 different operations:
- broadcast
- reduce
- scatter
- gather
- allgather
- reduce-scatter
- allreduce
Each operation has three different implementations:
- 'short' optimised for short vectors
- 'long' optimised for long vectors
- 'builtin' which exclusively uses built-in MPI operations (calls to
MPI_Bcast,MPI_Scatter, etc)
The algorithms in this repository are given in the following paper:
Chan, E., Heimlich, M., Purkayastha, A. and van de Geijn, R. (2007), Collective communication: theory, practice, and experience. Concurrency Computat.: Pract. Exper., 19: 1749-1783. https://doi.org/10.1002/cpe.1206
The computing for this project was performed at the OU Supercomputing Center for Education & Research (OSCER) at the University of Oklahoma.
Consequently, the shell scripts to build the project are built for that system.