ENERGY-AWARE SCHEDULING: OPTIMIZING POWER EFFICIENCY IN EXASCALE COMPUTING ENVIRONMENTS

Abstract

Exascale systems have crossed a threshold that changes how we think about performance. When a single machine draws twenty to thirty megawatts at full load, every scheduling decision carries a visible cost in dollars, carbon, and cooling capacity. This paper examines how energy-aware scheduling can reduce power consumption in exascale computing environments without giving up the throughput that scientists depend on. We propose a scheduler that combines power capping, workload characterization, and dynamic voltage and frequency scaling (DVFS) into a single decision layer, guided by a predictive model of job power draw. Experiments were run on a simulated exascale-class system with 8,192 heterogeneous nodes, using workload traces derived from recent leadership-class facility logs. Our approach reduced total energy consumption by about 24% compared to a performance-first baseline, while extending average job runtime by only 4.6%. Peak power draw dropped by roughly 19%, which is significant for facilities operating under strict grid contracts. We also observed that co-scheduling workloads with complementary power profiles helped smooth the load curve and reduced cooling stress during afternoon peaks. Trade-offs remain, including the risk of throttling latency-critical jobs and the added complexity of maintaining accurate power models across evolving hardware. Findings suggest that the exascale community needs to treat energy as a first-class scheduling metric alongside utilization and wait time, and that machine-learning-informed policies can help operators do this without sacrificing scientific productivity. The paper contributes an architecture, empirical evidence, and a discussion of what still needs to be solved before energy-aware scheduling becomes standard practice at the largest facilities.