Performance of an OO compute kernel on the JVM: revisiting Java as a language for scientific computing applications

The study of Java as a programming language for scientific computing is warranted by simpler, more extensible and more easily maintainable code. Previous work on refactoring a C++ scientific computing code base to follow best practises of object-oriented software development revealed a coupling of such practises and considerable slowdowns due to indirections introduced by abstractions. In this paper, we explore how Java's JIT compiler handle such abstraction-induced indirection using a typical scientific computing compute kernel extracted from a linear solver written in C++. We find that the computation times for large workloads on one machine can be on-pair for C++ and Java. However, for distributed computations, a better parallelisation strategy needs to be found for non-blocking communication. We also report on the impact on performance for common "gripes": garbage collection, array bounds checking, and dynamic binding.