This is awesome. Although it looks like the parallelism is set up in terms of partitioning a large alignment, rather than partitioning a large number of taxa. I could be wrong though.
We are indeed partitioning/parallelizing over the alignment, doing so over the taxa is very difficult.
The paper will be submitted in the next 10 days or so.
@Alexis_RAxML are we to understand that functionally this is going to be very similar to MrBayes? Did you follow the MB code at all during development?
@ematsen yes, it implements a similar set of proposals as MrBayes (e.g., the excellent parsimony-guided SPR move from Lakner2009). It uses the RAxML likelihood implementation and adds three levels of parallelism (MPI), such that several computing nodes can be used to carry out large-scale analyses.