Performance Characterization and Evaluation of HPC Algorithms on Dissimilar Multicore Architectures

Abstract

In this paper, we share our experiences in using two important yet different High Performance Computing (HPC)architectures for evaluating two HPC algorithms. The first architecture is an Intel x64 ISA based homogenous multicore with Uniform Memory Access (UMA) type shared-memory based Symmetric Multi-Processing system. The second architecture is an IBM Power ISA based heterogenous multicore with Non-Uniform Memory Access (NUMA) based distributed-memoryAsymmetric Multi-Processing system. The two HPC algorithms are for predicting biological molecular structures, specifically the RNA secondary structures. The first algorithm that we created is a parallelized version of a popular serial RNA secondary structure prediction algorithm called PKNOTS. The second algorithm is a new parallel-by-design algorithm that we have developed called MARSs. Using real Ribo-Nucleic Acid(RNA) sequences, we conducted large-scale experiments involving hundreds of sequences using the above two algorithms. Based on thousands of data points that we collected as an outcome of our experiments, we report on the observed performance metrics for both the algorithms on the two architectures. Through our experiments, we infer that architectures with specialized coprocessors for number-crunching along with high-speed memory bus and dedicated bus controllers generally perform better than general-purpose multi-processor architectures. In addition, we observed that algorithms that are intrinsically parallelized by design are able to scale & perform better by taking advantage of the underlying parallel architecture. We further share best practices on handling scalability aspects with regards to workload size. We believe our results are applicable to other HPC applications on similar HPC architectures.