Affordable overhead in CORRELATE: Comparing two MD simulators

Abstract

In this paper, we illustrate the way HPC application programs are developed using CORRELATE, a concurrent object-oriented language that supports applications for distributed memory systems. Our programming environment favours HPC applications because CORRELATE is a heterogeneous language that extends C++: on the one hand, CORRELATE objects can be coarse grained entities that exploit the language support because they model the concurrency aspects that are inherent to parallelism. On the other hand, each CORRELATE object can encapsulate C++ entities to implement its behaviour.

CORRELATE is a truly object-based language in that it enforces the application programmer to model an application as a collection of objects that interact by invoking operations on each other. Even though the development of commercial software is more and more dominated by the use of object-oriented technology for obvious reasons (ensuring that development efforts can be paid back in real sales), HPC applications are seldom built using OO language technology because of the potential overhead that is generated by an advanced object support system. The CORRELATE programming system attempts to deliver the best of both worlds by supporting concurrent objects as well as sequential objects. The latter can be implemented efficiently in C++, and concurrent objects simplify the modelling of interactions and synchronisation between high level application objects.

These aspects are illustrated by showing two CORRELATE implementations of an application in the area of molecular dynamics. The first example uses a coarse-grained approach in which subspaces of the simulated geometry are modelled as CORRELATE objects — and in that sense, these objects look similar to individual processes in SPMD programs. As an extreme, the second application models each particle as a CORRELATE object.