Computer Science Department Events
The Computer Science Department will be hosting a colloquium on Friday, December 7, 2012 in the E&CS Building first floor auditorium.TIME: 10:30 (DONUTS) 10:40 (TALK)Hybrid Parallelism and Memory Affinity for Large Shared Arrays with Applications to Nuclear Structure CalculationsDr. Masha Sosonkina
E & C S BUILDING AUDITORIUM (1ST FLOOR)
Department of Modeling, Simulation, and Visualization Engineering
Old Dominion UniversityAbstract:
With the increase in the processing core counts on modern computing platforms, main-memory accesses present a considerable execution bottleneck, leading to poor scalability in multithreaded applications. Even when the memory is physically divided into separate banks, each associated with a set of cores, i.e., exhibiting the so-called nonuniform memory access (NUMA) architecture, the access time to the shared data structures may be detrimental to performance. Hence, it is imperative to carefully map large shared arrays to specific memory banks based on the nature of the computation and characteristics of the multithreaded parallelism. In this talk, we describe memory-pinning strategies, applicable to such typical numerical kernels as sparse matrix-vector multiplication and vector orthogonalization, aiming to minimize remote memory accesses. Experiments with parallel benchmarks as well as with a realistic application calculating ab initio nuclear structure will be shown. Performance gains of up to 3.5 times were observed with the proposed strategies compared with the default memory placement policy on the Cray XE6 architecture. Several nuclei and nuclear interactions were considered in the large-scale test cases with the dimensions of the sparse symmetric matrices ranging from one million to one billion. We also compare the performance of the pure message-passing (MPI) versions with the hybrid MPI/multithreaded ones.