Measuring and Improving Application Performance with PerfSuite
This article has touched only the surface of the techniques available to you when using hardware performance counters to measure and improve the performance of your applications. Hopefully, you now have an idea of what hardware performance counters are and how they can help you gain insight into performance bottlenecks. If you would like to get started using PerfSuite or other tools and supporting software mentioned in this article, visit the on-line Resources.
Many different ways exist in which applications can be tuned for higher performance. In fact, the most effective way is not loop-level improvements or tweaking but fundamental changes to the algorithms used in your application that are more computationally efficient. Ideally, your software will use efficient algorithms further tuned to make effective use of your CPU. PerfSuite and other similar tools can go a long way toward making this process easier for you.
I would like to thank Professor Danesh Tafti of the Mechanical Engineering Department at Virginia Tech for providing the program used for the psrun profiling example in Listing 5. This is a computational kernel extracted from a computational fluid dynamics application named GenIDLEST that Tafti and his research team use, maintain and develop. I also would like to express my thanks to all the PAPI team members of the Innovative Computing Laboratory at the University of Tennessee-Knoxville for their support and encouragement during the development of PerfSuite.
Resources for this article: /article/8264.
Rick Kufrin currently is a senior member of the technical staff at the University of Illinois' National Center for Supercomputing Applications. He is the originator and technical lead for the PerfSuite software project described in this article and is available for consultation on the use of PerfSuite and other technologies for software improvement.