Thread-based parallelization

[stamatak]

I am just working on developing a PThreads-version of ExaML.

In this context I observed the following:

1. We mainly need to implement a function analogous to the MPI allreduce operation in PThreads.
2. Thinking of auto-tuned likelihood implementations, we may also have to consider aut-tuned barrier implementations. The reason is that barrier performance and types required will probably be very different if we just have one partition or several partitions that are mostly distributed monolithically to threads.

This could be a nice technical paper in the context of a master's thesis.

[xflouris]

After the white-board session on 6.10.2015 we decided to implement an autistic opt.and a collaborative opt. and test it again the MPI version in ExaML.

As a first step we should compare them using 16 cores on a dataset with 16 partitions and 1000 sites.