High performance computing systems are moving towards the exaflops era. The tremendous increase in computational speed is accompanied by enormous power consumption in these systems. It is necessary to harvest any potential opportunities to save power in these high end computing systems. The goal of this paper is to explore possibilities of power savings in the interconnects between the nodes. By careful scheduling of jobs in a 3D torus-connected cluster of nodes, we show that significant amounts of power can be saved by switching certain portions of the network elements to low power modes. We also present an estimation method that more accurately estimates the actual runtime of jobs from the user provided runtimes and enhances the performance of the scheduling scheme. We validate our results via detailed MATLAB simulations.
Copyright 2013 Elsevier B.V., All rights reserved.
- Energy efficiency
- Extensible bin packing
- Green HPC
- HPC scheduling