Hybrid MPI/OpenMP programming on the Tilera manycore architecture

This work assesses the viability of different programming models for large-scale manycores using an MPI-like abstraction, the vendor's OpenMP, and a combination (hybrid) of both. Experiments with Tilera's TilePro64 demonstrate that MPI and OpenMP both scale while the hybrid model performs inferior to the others. Further, network-on-chip contention significantly affects performance and its variance, especially if the number of utilized cores is high. These findings provide an early insight on trends for single die/chip solutions with large numbers of cores, which will become mainstream in HPC in the coming years. Prior work lacks such a study on large manycores with an efficient native on-chip message passing on shared memory on the same hardware platform.