More for Less: Integrating Capability-Predominant and Capacity-Predominant Computing
Zhong Zheng, Michael E. Papka, Zhiling Lan
TL;DR
This paper tackles resource underutilization and cost inefficiency arising from siloed capability-predominant and capacity-predominant HPC platforms. It conducts a trace-based, event-driven analysis using an extended CQSim simulator on real Theta (capability-predominant) and Cori (capacity-predominant) workloads to evaluate two unification strategies: workload fusion, which co-schedules both workload types on a single unified system, and workload injection, which backfills a capability system with capacity jobs. The study shows that workload fusion can boost overall resource utilization and reduce wait times, with potential cost savings from downsizing the unified system; workload injection can further improve Theta's utilization by filling temporal/spatial gaps with Cori's short/small jobs, while incurring manageable impacts on Theta's job wait times. The findings offer practical guidance for designing unified HPC infrastructures and motivate open-sourcing the simulation tools and data for broader community use.
Abstract
Capability jobs (e.g., large, long-running tasks) and capacity jobs (e.g., small, short-running tasks) are two common types of workloads in high-performance computing (HPC). Different HPC systems are typically deployed to handle distinct computing workloads. For example, Theta at the Argonne Leadership Computing Facility (ALCF) primarily serves capability jobs, while Cori at the National Energy Research Scientific Computing Center (NERSC) predominantly handles capacity workloads. However, this segregation often leads to inefficient resource utilization and higher costs due to the need for operating separate computing platforms. This work examines what-if scenarios for integrating siloed platforms. Specifically, we collect and characterize two real workloads from production systems at DOE laboratories, representing capabilitypredominant and capacity-predominant computing, respectively. We investigate two approaches to unification. Workload fusion explores how efficiently resources are utilized when a unified system accommodates diverse workloads, whereas workload injection identifies opportunities to enhance resource utilization on capability computing systems by leveraging capacity jobs. Finally, through extensive trace-based, event-driven simulations, we explore the potential benefits of co-scheduling both types of jobs on a unified system to enhance resource utilization and reduce costs, offering new insights for future research in unified computing.