Computer Science · Hardware and Architecture
Exploiting Object-Oriented Abstractions to parallelize Sparse Linear Algebra Codes
Christian Terboven, An Mey Dieter, Paul Kapinos, Christopher Schleiden, Igor Merkulow
See on the mapAbstract
Object-oriented (OO) programming conquered mainstream in the 1990s but the procedural programming languages Fortran and C are still dominating in the High-Performance-Computing (HPC) community. However in recent years, compute kernels or even whole applications have been migrated from C to object-oriented C++ and Fortran has been retrofitted with OO capabilities as well. In this work we illustrate how OO abstractions can be used to introduce Shared-Memory parallelization into existing applications by using the concepts of encapsulation and modularity. For the important domain of sparse linear algebra routines, we discuss several parallel implementation strategies and data type design considerations. We aim to maximize performance and scalability while hiding the implementation details as much as possible.