Intel Paragon MPP: getting the most from all of the nodes

Abstract

The intel Paragon may no longer be considered state-of-the-art technology in terms of performance capabilities, but for a hybrid GP/MP system such as the one at RUS, it is possible to 'stretch' the performance of the system by utilizing the dual-CPUs of the MP nodes at the same time as the single-CPU GP nodes. The Paragon at RUS can then calculate loops using up to 154 CPUs vs. the 113 compute CPUs which would otherwise be used. The trick is balancing the load; a process for which, in this case, no standard library exists.