libPowerMon: A Lightweight Profiling Framework to Profile Program Context and System-level Metrics

摘要

As power becomes one of the most important resources to provision while building modern HPC systems and applications, it becomes crucial to obtain deeper insights into applications' power and thermal characteristics. There exists a need to correlate application context with processor-level and system-level power and thermal measurements. Existing profiling tools to monitor power and thermal measurements either operate at a granularity that is not fine enough to correlate with application-level events that describe application context or are not equipped to sample application-level events. In this work, we introduce libPowerMon, a lightweight user-level profiling framework to simultaneously sample user-specified application events and system-level metrics at up to 1 kHz sampling intervals. At the application level, libPowerMon provides a source-level phase markup interface to capture application context. It records MPI and OpenMP events, and samples processor state at a finer temporal granularity. At the system level, libPowerMon samples power and thermal characteristics and provides an interface to set processor and DRAM power. We present three case studies that demonstrate the benefits of libPowerMon in saving cluster-level power and improving application performance within a system-enforced power limit.