Adaptive Sampling of Performance Counters

Many applications of profiling based on sampling of Performance Counters (PC), such as feedback-directed optimization and software reliability, are often constrained by the amount of information that can be obtained without perturbing significantly the behavior of the profiled task. Current implementation of event and time based sampling software utilize fixed or random sampling periods which are unresponsive to changes on the frequency of occurrence of architectural events. In this paper we introduce an adaptive sampling schema for event-based sampling of PCs that adaptively varies the sampling period in order to minimize the reconstruction error and reduce the number of sample points taken. By restricting the number of samples taken, AES prevents oversampling and overload to the profiled task. We conducted experiments within the perf event kernel module and have observed a decreased reconstruction error (MSE) with respect with the ubicuos fixedperiod sampling. An additional objective of this work is to motivate the formal study of the properties of sampling Hardware Performance Counters, an optimization opportunity that has rarely been explored up to now within this context.