Hybrid MPI/Pthread Implementation of 1-D FFT on SMP
Rami Al Na'mneh, W. David Pan · 2006
SMP (shared memory multiprocessors) can support two programming models: message passing and shared address space. Shard address space offers easy programming, particularly for irregularly structured computations, but suffers from performance limitation due to the protocol overhead and poor spatial locality (false sharing). On the other hand, it has been shown that there are performance advantages in integrating message passing in cache coherent multiprocessors. In this paper, we propose the hybrid MPI/Pthread implementation of 1-D FFT on SMP. We compare the results of hybrid MPI/Pthread implementation with pure MPI implementation. The results show that MPI/Pthread implementation perform better than pure MPI implementation for large data sizes.