Process Variation-Aware Power Scheduling for HPC Applications
Student: Neha Gholkar (North Carolina State University)
Supervisor: Frank Mueller (North Carolina State University)
Abstract: Until recently research community has focused on minimizing energy usage
of super computers.
Considering the US DoE's mandate of power constraint of 20 MW for the exascale sites,
efforts need to be directed towards minimizing the wasteful
usage of power while maximizing performance under this constraint.
Most of the workloads on supercomputers are often coupled parallel
scientific simulations. It has been observed that
processors tend to be non-uniform with respect to their performance under power bounds. We also observe that the process variability translates into variation in peak power efficiency of the processors.
Hence, uniform power capping, a naive strategy of enforcing a job level power budget, does not lead to optimal solutions.
We propose a variation-aware algorithm that optimizes a job for performance
under a power constraint. Our experiments show that our algorithm achieves up to 29% improvement in performance as compared to uniform power capping.
Two-page extended abstract: pdf