Automated techniques for mapping CSP-based concurrent programs to compute clusters

Abstract

Distributing an application's processes across a cluster of nodes is a viable method for improving computational performance, therefore this lead to researching a number of algorithms which use different techniques to map applications onto the underlying architecture. This dissertation seeks to create this same mapping effect on CSP-based concurrent applications in order to evaluate how different mapping algorithms will perform in such cases. A CSP library was first developed to provide the necessary tools and functionality for implementing CSP-based concurrent applications. It enables seamless communication between processes on the same node and processes across different nodes. The purpose of this library is to be able to design a number of CSP-based applications with different parallel patterns. A new configuration language was implemented to provide a straightforward way to map such CSP-based applications onto an available cluster. This was designed in such a way to allow for mapping the same application using different mapping algorithms without having to recompile the application. This framework developed for this dissertation will save application developers a signifi cant amount of time and effort to generate mappings for their individual CSP applications. This work was then used to compare well-known mapping algorithms to see how they affect the performance of CSP-based applications when mapped onto compute clusters. After evaluating each of the mappings generated for the different applications, results show that even when dealing with CSP-based applications, certain mappings generate close to optimal performance for certain applications, whereas the same mappings result in poor performance when applied to other applications. Results indicate that application performance is affected by the number of channels introduced between partitions, as shown in the case of application graphs with a large tree-depth. Moreover, distributing processes unevenly, especially in the case of application graphs with a short tree-depth, also significantly affects an application's performance.