Fine grained parallel computation in the cloud

Abstract

The divergence of priorities between high performance computing (HPC) and cloud infrastructure has made scaling tightly coupled parallel applications in the cloud less viable than their supercomputer counterparts. As a result, many potential benefits -- elasticity, job virtualization, and cost-effectiveness -- of cloud computing remain underutilized by the HPC community. Through analysis and benchmarking of cloud compute services we see that network overhead on cloud applications leads to scalability issues in fine grained parallel computations. Our approaches seek to improve scalability of a parallel runtime system and develop new runtime system methods to hide latency within cloud scale applications. To improve scalability of communication within a node, we developed a new concurrent lock-free queue that supports a large memory bound and efficiently reclaims memory without a need for an external memory reclamation scheme. Additionally, we implement a replication system for parallel compute objects to reduce latency of requests.