CosmoCoffee

Hello,
It seems not true that the more CPUs the better to run CosmoMC on a cluster. How many CPUs is the most efficient? Thank you very much for suggestions.

I recommend running about 6 chains (so 6 MPI nodes), with up to 4 processors per node (openmp). I think the WMAP3 code is not really parallelized, so probably doesn't help very much having more processors per node. Personally I usually run 6x2.

But surely, it spends most of the time inverting matrices, so if you use MKL lapack and set OMP_NUM_THREADS to 4 it should be quite parallelised? (but in practice I've noticed suboptimal use of processors, which I couldn't explain...)