Performance engineering of hybrid message passing + shared memory programming on multi-core clusters

The hybrid message passing + shared memory programming model combines two parallel programming styles within the same application in an effort to improve the performance and efficiency of parallel codes on modern multi-core clusters. This thesis presents a performance study of this model as it applies to two Molecular Dynamics (MD) applications. Both a large scale production MD code and a smaller scale example MD code have been adapted from existing message passing versions by adding shared memory parallelism to create hybrid message passing + shared memory applications. The performance of these hybrid applications has been investigated on different multi-core clusters and compared with the original pure message passing codes. This performance analysis reveals that the hybrid message passing + shared memory model provides performance improvements under some conditions, while the pure message passing model provides better performance in others. Typically, when running on small numbers of cores the pure message passing model provides better performance than the hybrid message passing + shared memory model, as hybrid performance suffers due to increased overheads from the use of shared memory constructs. However, when running on large numbers of cores the hybrid model performs better as these shared memory overheads are minimised while the pure message passing code suffers from increased communication overhead. These results depend on the interconnect used. Hybrid message passing + shared memory molecular dynamics codes are shown to exhibit different communication profiles from their pure message passing versions and this is revealed to be a large factor in the performance difference between pure message passing and hybrid message passing + shared memory codes. An extension of this result shows that the choice of interconnection fabric used in a multi-core cluster has a large impact on the performance difference between the pure message passing and the hybrid code. The factors affecting the performance of the applications have been analytically examined in an effort to describe, generalise and predict the performance of both the pure message passing and hybrid message passing + shared memory codes