Summary and Conclusions

We discussed the use of hardware instrumentation mechanisms that are universally available on modern processors and chipsets as a basis for a performance reflection facility. Using this facility, it possible to estimate productivity and overhead metrics. Systems software can use these two metrics to improve its performance. We showed several potential uses: The OS kernel can use the metrics to tune its scheduling decisions. Applications can use these metrics to determine the concurrency level. Finally, we provided a working example for using reflection in server throttling to prevent thrashing.