Free registration here
Abstract: n the era of Cloud Computing, Big Data, and Quantum Physics Simulations, data centers play a key role in the world ICT infrastructure. The need to match the ever-growing demand of computing services within a reasonable power envelope has pushed modern data centers to adopt techniques to limit their energy requirements, such as using heterogenous architectures, free cooling, or virtualization. While effective at mitigating the consumption, such techniques significantly increase the complexity of workload dispatching, giving a hard time to the existing scheduling systems. In this context, Combinatorial Optimization methods have the opportunity to be the enabling factor for the next generation of job dispatchers in data centers.
In this work, we focus on the problem of job dispatching on a real supercomputer (for High Performance Computing applications) having heterogeneous architecture, namely the EURORA system installed in the CINECA data center in Bologna. The problem consists in mapping and scheduling a stream of computation-intensive jobs with approximately known duration on the supercomputer resources. Currently, this is mostly done done via a rule-based system (Altair PBS) that incurs the risk of causing resource fragmentation (and hence underutilization) or large waiting times.
We are working on an alternative approaches based on Constraint Programming and Combinatorial Optimization in general. A series of prototypes has already been realized and deployed both in a simulated environment and on the real supercomputer. The approach is leading to significant improvements in terms of waiting times and comparable machine utilization when compared to the system currently installed in the data center.
Biography: Michele Lombardi is a an assistant professor (no tenure track) at University of Bologna. He is working on the integration of heterogeneous techniques for Combinatorial Optimization, and on hybrid off-line/on-line optimization. His expertise is on Constraint Programming, Integer Linear Programming and Machine Learning, with main applications on resource allocation and scheduling problems. Michele has a PhD in Electrical, Computer and Telecommunications Engineering, from University of Bologna. He received the AI*IA "Marco Cadoli" PhD award in 2010, and honorable mentions at the CP 2011 and ICAPS 2012 PhD awards.