Evaluation of Parallel I/O Performance and Energy with Frequency Scaling on Cray XC30

Abstract—Large-scale simulations produce massive data that needs to be stored on parallel file systems. The simulations use parallel I/O to write data into file systems, such as Lustre. Since writing data to disks is often a synchronous operation, the application-level computing workload on CPU cores is minimal during I/O and hence we consider whether energy may be saved by keeping the cores in lower power states. To examine this postulation, we have conducted a thorough evaluation of energy consumption and performance of various I/O kernels from real simulations on a Cray XC30 supercomputer, named Edison, at the National Energy Research Supercomputing Center (NERSC). To adjust CPU power consumption, we use the frequency scaling capabilities provided by the Cray power management and monitoring tools. In this paper, we present our initial observations that when the I/O load is high enough to saturate the capability of the filesystem, down-scaling the CPU frequency on compute nodes reduces energy consumption without diminishing I/O performance. I