Compilation Guidelines

Juropa and other Intel machines

--with-alignment

A number of benchmarks on JUROPA suggests that when compiling with ICC, the alignment setting makes essentially no difference. The only exception is a value of 16, which gives a mean with the lowest variance but all the values overlap within errors.

--with-mpidimension

For 8 nodes (the maximum interactive job size on Juropa) there is no apparent statistically significant difference between different parallelisations. This should be checked with more nodes as the fat tree runs out of bandwidth.

--with-persistentmpi

On Juropa, at least with only 8 nodes (64 MPI tasks), this leads to substantial gains up to 10% in the benchmark with communication when compared to the default.

JuGene

JuQueen / BG/Q Cineca

Shm / CRAY & co.

small clusters with slow interconnects