Just how accurate are performance counters?

The wide use of performance counters by application developers and benchmarking teams gives evidence that performance counters are well worth the silicon and design time required to include them on modern microprocessors. These counters provide rudimentary performance measurements that may or may not be accurate. This paper presents our methodology for determining the accuracy of these counters as well as preliminary results of a study that, using this methodology, evaluates the accuracy of the R12000 performance counters with respect to eight of 30 measurable events. The results indicate that care must be taken when using data generated by performance counters because, in some cases, this data may lead to erroneous conclusions. This can occur when the granularity of the measured code is not sufficient to ensure that the overhead introduced by counter interfaces does not dominate the event counts.