Fully Predictable HPC Infrastructure using Admission Control with Virtualization

Historically, batch scheduling has dominated in managing HPC workloads despite its unpredictability regarding job’s wait time. Although existing researches such as reservation partially solved the problem, fully predictable HPC system still remains elusive goal while emerging adaptive applications urge its realization. Our earlier study presented a control-theoretic, VM-based approach that achieves good predictability. However, it was a part of complete solution as it does not scale to systems with high workloads. This paper presents a design and implementation of fully predictable HPC system using admission control (AC) as predictability semantic. The system makes concrete admission decision regarding a job’s deadline, even when the system is overloaded by unpredictable workloads. Our approach is based on utilization test and adaptive control as theoretical underpinnings of predictability, and use VM’s checkpointing and online CPU reconfiguration. Our experimental results using 5 widely used applications show the AC-enabled HPC system achieves high predictability with no significant overhead.