PANN: Power Allocation via Neural Networks Dynamic Bounded-Power Allocation in High Performance Computing

Abstract

Exascale architecture computers will be limited not only by hardware but also by power consumption. In these bounded power situations, a system can deliver better results by overprovisioning -having more hardware than can be fully powered. Overprovisioned systems require power to be an integral part of any scheduling algorithm. This paper introduces a system called PANN that uses neural networks to dynamically allocate power in overprovisioned systems. Traces of applications are used to train a neural network power controller, which is then used as an online power allocation system. Simulation results were obtained on traces of ParaDiS and work is continuing on more applications. We found in simulations PANN completes jobs up to 24% faster than static allocation. For tightly constrained systems PANN performs 6% to 11% better than Conductor. A runtime system has been constructed, but it is not yet performing as expected, reasons for this are explored.