Energy-Efficient Planning for Heterogeneous Computing Systems
Nowadays, most computing centers are built by gathering a heterogeneous collection of computers. In fact, heterogeneous computing (HC) systems have notoriously increased in the last twenty years, both in traditional cluster infrastructures and in modern grid computing systems, which are conceived as a large loosely-coupled virtual supercomputer formed by combining many heterogeneous platforms of different characteristics. Clusters and grid infrastructures have made it feasible to provide pervasive and cost-effective access to computing resources for solving hard problems, an approach that grid computing pioneers Ian Foster and Carl Kesselman have covered in detail.
A key problem when using HC systems consists in finding an optimal scheduling strategy for a set of tasks to be executed. The goal is to assign tasks to the computing resources by satisfying some efficiency criteria. The scheduling problem usually implies several conflicting objectives to optimize, such as those matching the interests of the service provider (total execution time, infrastructure utilization, cost, etc.), and the customer, who is especially concerned with Quality of Service (QoS). Practitioners in HC scheduling are searching for suitable implementations of effective and efficient algorithms. In this context, heuristics are useful methods that rely on simple optimization strategies, thus they are able to compute accurate schedules in reasonable time.
Recently, power provisioning and energy consumption have become great challenges in the field of computing. Energy consumption is now of major interest, raising various monetary, environmental, and system performance concerns. Energy consumption on computing systems is not only dependent on the energy efficiency of the hardware, but also on the infrastructure management software such as the resource management system and the workload scheduling algorithms.
Thus, researchers have focused on developing energy-aware scheduling algorithms for HC systems. Service providers are always interested in more energy-efficient solutions with reasonable QoS level. Although energy optimization has been tackled at the computing element level, showing to be effective, energy efficiency at the cluster/grid level is of recent interest. The main trend is to optimize the energy consumption of the the computing elements, because the processor is the main consumer of energy, and it also offers the most flexible options for energy management, such as dynamic voltage scaling (DVS), dynamic power management, slack sharing and reclamation, and other techniques.
Our recent work “Energy-Aware Scheduling on Multicore Heterogeneous Grid Computing Systems” (by S. Nesmachnow, B. Dorronsoro, J. Pecero, and P. Bouvry, published in Journal of Grid Computing, Springer), tackles the energy-aware scheduling problem in HC systems. We targeted the design of twenty highly efficient multi-objective heuristics that provide different trade-off solutions to the problem for multicore HC and grid systems, minimizing both the execution time (makespan) and the energy consumption.
We proposed an approach based on concepts from state-of-the-art heuristics for makespan optimization, but applying a strategy to take into account the energy consumption objective. We use the two-phase structure of traditional list scheduling heuristics and apply diverse combinations of execution time and energy optimization, to account for both objectives simultaneously. This simple methodology allows building novel heuristics that provide one single accurate trade-off solution in a quick response, and they are suitable for being implemented in nowadays real systems.
Our optimization approach does not apply specific power management techniques. Instead, we propose a mathematical model for explicitly computing the energy consumption in multicore systems relying on the Max-Min mode and a linear approximation for the energy consumed when using different number of cores. This model accounts for hardware-embedded energy saving features, such as SpeedStep technology by Intel, or Optimized Power Management by AMD. By using our explicit model, the proposed heuristics perform a two-dimension energy optimization: within a computer (using multiple cores) and across different HC/grid nodes.
The twenty fast multi-objective heuristics proposed in our article were evaluated over benchmark HC/grid problem instances built following a methodological approach based on well-known models for execution time estimation in HC systems. In addition, we used real up-to-date performance and energy consumption values for a set of 64 modern processors, gathered from specialized reports and websites. A scalability analysis was also carried out in order to study the efficacy of the best heuristics when facing scenarios with different ratios between the number of tasks and machines. As a result of the experimental evaluation, we demonstrate that by using a strategy combining the execution time and energy consumption objectives, schedules with a reduction of 6 to 14% in the energy consumption objective can be found, while the makespan objective is only increased by 2 to 6%. We have also shown that these results can be improved by applying simple multiobjective local search techniques.
Our work has provided a first step towards designing efficient energy-aware scheduling heuristics that can be easily integrated into modern middleware and management systems for HC facilities, such as the TORQUE/Maui manager and the OAR batch scheduler. We have shown that accounting for a single criterion is clearly not the best strategy when tackling the energy-efficiency planning on heterogeneous distributed computing centers. Both objectives, energy consumption and QoS, should be taken into account, as the different strategies proposed compute different trade-off results for the energy-aware scheduling problem.
About the Author
Sergio Nesmachnow has a Ph.D (2010), a M.Sc. (2004) in Computer Science, and a degree in Engineering (2000) from Universidad de la República, Uruguay. He is currently a Full Time Professor at Numerical Computing Center, Engineering Faculty, Universidad de la República, Uruguay, with several teaching and research duties. His main research areas are scientific computing, high performance computing, and parallel metaheuristics applied to solve complex real-world problems. He has published over 90 papers in international journals and conference proceedings. S. Nesmachnow serves as Editor-in-Chief for International Journal of Metaheuristics, and is Guest Editor in Cluster Computing and The Computer Journal. He also participates in several technical program committees of international conferences and is a reviewer for many journals and conferences. E-mail: sergion@fing.edu.uy, Webpage: www.fing.edu.uy/~sergion.
