Improving Reliability & Efficiency Of Performance Monitoring In Linux

Abstract

Processor hardware performance counters have improved in quality and features in recent years. At the same time, the performance monitoring support in Linux has been significantly revamped with the development of the perf_events subsystem. Those factors concur in making performance monitoring a more common practice among developers. However, no performance analysis is possible without reliable hardware counter data.In this thesis, we focus on a published correctness erratum in the performance monitoring unit of recent Intel processors when Hyper-Threading is enabled. This erratum causes cross hyper-thread hardware counter corruption and may produce unreliable results. We propose a cache-coherence style protocol that we implement in the Linux kernel to circumvent the issue by introducing cross hyper-thread dynamic event scheduling. We also introduce an event scheduling algorithm that achieves the optimal scheduling of events onto hardware counters at all times. The proposed optimizations do not require any user level changes and leverage the internal design of the perf_events subsystem.The source code has been posted to the upstream Linux kernel and our contribution has received honorary awards from Intel and Google.