Communication-Centric Solution for Cholesky Decomposition: Industrial Case Study for Coordination Programming

This paper is a performance study of a parallel programming model called S-Net. The purpose of S-Net is to support writing coordination code that instantiates components as "boxes" and connects them with anonymous data streams. An application is represented as a network between the standard input and output, which are two external streams connecting the whole application with its environment. We evaluate S-Net using a numerical algorithm for a linear algebra problem - Cholesky decomposition. Cholesky decomposition has many important applications. It is also provided as one of the sample applications for the Intel Concurrent Collections (CnC). Our measurements demonstrate that S-Net has a similar performance to CnC on this problem. For both implementations an effective utilisation of resources is achieved without platform-specific tuning. We compare them with the serial implementation and demonstrate the significant speedup for no programmability cost.