Computationally Efficient Multiplexing of Events on Hardware Counters

This paper proposes a novel approach for scheduling n performance monitoring events onto m hardware performance counters, where n > m. Whereas existing scheduling approaches overlook monitored task information, the proposed algorithm utilizes the monitored task’s behavior and schedules the combination of the most costly events. The proposed algorithm was implemented in Linux Perf Event subsystem in kernel space (build 3.11.3), which provides finer granularity and less system perturbation in event monitoring when compared to existing user space approaches. Benchmark experiments in PARSEC and SPLASH.2x suites compared the existing round-robin scheme with the proposed rateof-change approach. Results demonstrate that the rateof-change approach reduces the mean-squared error on average by 22%, confirming that the proposed methodology not only improves the accuracy of performance measurements read, but also makes scheduling multiple event measurements feasible with a limited number of hardware performance counters.