Proactive workload dispatching on the EURORA supercomputer

Abstract

In the era of Cloud Computing, Big Data, and Quantum Physics Simulations, data centers play in the world ICT infrastructure a role as big as (sadly) their power consumption. In many cases, a surprising amount of such consumption is due to idle resources, either introduced to face workload peaks or leftovers of workload fragmentation. In this context, proactive workload dispatching has the chance to improve the utilization of computing resources, thus reducing the idle time and improving the ability to handle peaks. In this paper, we devise a CP based approach for proactive workload dispatching on the EURORA supercomputer placed at the CINECA computing center in Bologna. The new system is evaluated on simulated job traces, where it leads to remarkable improvements in terms of both machine utilization and waiting times for queued jobs with respect to the currently used dispatcher, i.e., Portable Batch System (PBS). © 2014 Springer International Publishing Switzerland.