OpenMPI

OpenMPI / mpi4py

Manual on Multi processor usage

If you have OpenMPI and mpi4py installed, then you have access to Gary Thompson's multi-processor framework for MPI parallelisation.

However the code in relax must be written to support this. This is the case for the model-free analysis, in which case Gary has achieved near perfect scaling efficiency:

https://mail.gna.org/public/relax-devel/2007-05/msg00000.html

For the relaxation dispersion branch, no parallelisation has been attempted, neither in the original code from Sebastian Morin or the recent modifications by myself. This is not a simple task and will take a lot of effort to implement. If this is to be implemented one day, it is suggested parallelising at the level of the spin clusters.

It is often quite hard to achieve good scaling efficiency and often the first attempts will just make the code slower, even on a 1024 node cluster, due to the bottleneck of data transfer between the nodes.

The parallelisation will also require 10 times as much code to be written to do the same thing as non-parallised code, and debugging is much more difficult.

Update 2013/09/11

See Commit

Huge speed win for the relaxation dispersion analysis - optimisation now uses the multi-processor.

The relaxation dispersion optimisation has been parallelised at the level of the spin clustering.
It uses Gary Thompson's multi-processor framework. This allows the code to run on multi-core, multi -processor systems, clusters, grids, and anywhere the OpenMPI protocol is available.

Because the parallelisation is at the cluster level there are some situations, whereby instead of optimisation being faster when running on multiple slaves, the optimisation will be slower.
This is the case when all spins being studied in clustered into a small number of clusters.
It is also likely to be slower for the minimise user function when no clustering is defined, due to the overhead costs of data transfer (but for the numeric models, in this case there will be a clear win).

The two situations where there will be a huge performance win is the grid_search user function when no clustering is defined and the Monte Carlo simulations for error analysis.