HyperLoom Possibilities for Executing Scientific Workflows on the Cloud

摘要

We have developed HyperLoom - a platform for defining and executing scientific workflows in large-scale HPC systems. The computational tasks in such workflows often have non-trivial dependency patterns, unknown execution time and unknown sizes of generated outputs. HyperLoom enables to efficiently execute the workflows respecting task requirements and cluster resources agnostically to the shape or size of the workflow. Although HPC infrastructures provide an unbeatable performance, they may be unavailable or too expensive especially for small to medium workloads. Moreover, for some workloads, due to HPCs not very flexible resource allocation policy, the system energy efficiency may not be optimal at some stages of the execution. In contrast, current public cloud providers such as Amazon, Google or Exoscale allow users a comfortable and elastic way of deploying, scaling and disposing a virtualized cluster of almost any size. In this paper, we describe HyperLoom virtualization and evaluate its performance in a virtualized environment using workflows of various shapes and sizes. Finally, we discuss the Hyperloom potential for its expansion to cloud environments.