gbatch: Sustainable HPC Clusters via Renewable Energy–Aware Scheduling

Abstract

High performance computing (HPC) batch scheduling presents a unique opportunity to reduce operational carbon emissions through maximizing energy usage of data centers during periods of low-carbon grid intensity. At the same time, temporally shifting energy usage could cause a degraded user experience due to longer job delays. This thesis investigates two methods of green energy scheduling: a convex optimization technique and a greedy approach. Both approaches are evaluated on expected renewable energy usage and user wait time using simulations of real, six-day GPU job logs and associated power data collected from Princeton's High Performance Research Center. We find that a convex optimization scheduler that optimizes to maximize renewable energy usage and minimize user delay outperforms the greedy approach for both metrics. Although the convex optimization method increases the average user waiting time compared to the original schedule, it results in a 29.46% improvement in expected renewable energy use. We evaluate the effect of scheduling jobs using forecast renewable energy signals, instead of actual renewable energy signals, and determine that there is a 1% decrease in overall renewable energy utilization of the schedule generated using forecast renewable energy data compared to actual renewable energy data.