MPI and multi-threading in trans-abyss

[gwilymh]

Can anyone explain to me the difference between MPI processes and multithreading in trans-abyss?

Are the two mutually exclusive alternative forms of parallel processing? (The literature associated with abyss and trans-abyss was not entirely clear on this point)

[gwilymh]

A good answer to my questions is given here:
https://groups.google.com/forum/#!to...ss/_mHXGbtcO_U

MPI allows multi-threading between nodes in a cluster, while standard multi-threading can only utilize the CPU's within a node.

The very first stage of trans-abyss can use MPI in making de Bruijn graphs across multiple machines/nodes/motherboards. The latter stages of the analysis an only create threads amongst the CPU's of a single node.

"Suppose each machine on your computing cluster has 12 CPUs and 48 G of RAM. If your cluster support MPI and your cluster job is configured to use 2 machines, then you may run Trans-ABySS like so to make most use of the allocated resources:
transabyss --mpi 24 --threads 12 ...

Here, the 24 (12+12) MPI assembly processes can use up to 96 (48+48) G of RAM, whereas the non-MPI steps can only use up to 12 cpus and 48G of RAM. "