A Benchmark Suite for the OmpSs-2@ ArgoDSM task-based distributed shared memory environment

Abstract: Applications with high complexity, heavy computations and processing of large amount of data require powerful computer systems and programming models that allow efficient parallel execution without excessive programming effort. Benchmarking of these systems and models is of vital importance in order to identify their strengths, weaknesses, and potential improvements. The aim of this thesis is to create benchmarks for OmpSs-2@ArgoDSM, that is the combination of the OmpSs-2task-based parallel programming model with ArgoDSM, a reliable software solutionfor data coherence in distributed systems. In this effort, three benchmarks from PARSEC 3.0 suite are selected, implemented in parallel and measured initially using OmpSs-2 for SMPs, afterwards using OmpSs-2@Cluster for execution on distributed systems, and finally using OmpSs-2@ArgoDSM. Having implemented various versions of every benchmark, their performance and scaling is demonstrated, and a comparative study is conducted. The results indicate that, in general,OmpSs-2@ArgoDSM can achieve similar performance to that of OmpSs-2@Cluster,while the programs are significantly easier to implement. What is more, distributed execution using OmpSs-2 requires either zero data sharing,or serious changes in the code and compromise in the expected performance. Lastly,for applications that process regular size data sets, OmpSs-2@Cluster andOmpSs2@ArgoDSM can equally compete with OmpSs-2’s performance only when the data migration is extremely low and the partial processing workload is relatively heavy.