Parallelization

If GROOPS is compiled with the Message Passing Interface (MPI), most GROOPS programs can be run in parallel on multiple processor cores. Processing on computer clusters with distributed memory is also supported.

Many loops are parallelized by computing each loop step at a different core. Usually the first node distributes the work load, assigns loop steps to different cores, and is not participating on the actual loop computation. This means running GROOPS with only two nodes has no advantages in almost all cases. Non-parallel parts and programs without parallel support are executed at the first node only.

Large systems of normal equations, which are divided into blocks, are distributed over the nodes to reduce the memory consumption on each single node.

As all nodes may read and write files (at least reading the config files) the required part of the file system must be available on all participating computers.