Paper detail

Group Communication Patterns for High Performance Computing in Scala

We developed a Functional object-oriented Parallel framework (FooPar) for high-level high-performance computing in Scala. Central to this framework are Distributed Memory Parallel Data structures (DPDs), i.e., collections of data distributed in a shared nothing system together with parallel operations on these data. In this paper, we first present FooPar's architecture and the idea of DPDs and group communications. Then, we show how DPDs can be implemented elegantly and efficiently in Scala based on the Traversable/Builder pattern, unifying Functional and Object-Oriented Programming. We prove the correctness and safety of one communication algorithm and show how specification testing (via ScalaCheck) can be used to bridge the gap between proof and implementation. Furthermore, we show that the group communication operations of FooPar outperform those of the MPJ Express open source MPI-bindings for Java, both asymptotically and empirically. FooPar has already been shown to be capable of achieving close-to-optimal performance for dense matrix-matrix multiplication via JNI. In this article, we present results on a parallel implementation of the Floyd-Warshall algorithm in FooPar, achieving more than 94 % efficiency compared to the serial version on a cluster using 100 cores for matrices of dimension 38000 x 38000.