1 Reliability-aware Optimal K Node Allocation of Parallel Applications in Large Scale High Performance Computing Systems

The growing demand for more computational power to solve complex scientific problems is driving the physical scale of the system to hundreds and thousands of nodes. Ideally, scaling up the number of nodes should minimize the completion time, but algorithmic and system environment factors limit the scalability of parallel applications. Reliability is one of the major factors that limit performance, especially when applications are scaled to thousands of nodes. In this paper, we propose a reliability-aware optimal k-node allocation algorithm and compare it with Round Robin and reliability-aware resource allocation algorithms. Our simulation results indicate that giving flexibility to the resource manager in choosing the optimal number of nodes for large scale parallel applications based on the reliability of compute nodes minimizes the overall completion time and waste tim