Evaluation and analysis of a dynamic allocation scheme on a C++-based parallel programming library

We are currently studying the implementation of a C++-based parallel programming library to simplify parallel programming and efficiently execute parallel programs on a locally distributed computers with nonuniform performances. The class library employs dynamic allocation scheme of Light-Weight Processes (LWPs) and efficient execution of parallel applications. In this paper, we describe the dynamic allocation scheme implemented in the class library, discuss the experimental evaluation with parallel applications, and provide comparison with static allocation. The experiments results show that with dynamic allocation implemented in the class library, there is a significant performance advantage on a cluster environment with machines having nonuniform performances. Furthermore, this study provides insights on the trade-offs between task or process allocation and memory-access costs.