Using PerfExpert for parallel software development with Ashay Rane
Wednesday, September 7, 9:30 a.m., Computing Commons 120
Ashay Rane returns to ASU to present his work on PerfExpert, a profiling tool developed at the University of Texas at Austin. The goal of PerfExpert is to facilitate the performance debugging and tuning process for the users of high performance computing machines. PerfExpert automates the selection of performance counters, carries out multiple measurements of the code to gather the necessary data, diagnoses the causes of performance bottlenecks and recommends optimizations to correct the identified performance bottlenecks in plain English.
Performance tuning of parallel programs remains an art because of the complexity and ambiguity involved in three main steps:
- choosing the right metrics for measurement
- making sense of the performance metrics
- selecting the appropriate optimization(s) from a “knowledge base” of possible optimizations
Detailed knowledge of computer architecture is often required, which makes makes profiling and tuning programs especially difficult for users of high performance computing machines including chemists, geologists, physicists and others. As a result, many programs that run on current clusters often do not obtain maximum performance on the cores they utilize.
Ashay Rane graduated from Arizona State University in 2009 with a master’s degree in computer science. His thesis work on performance implications of hybrid MPI/OpenMP programming was done at the ASU Advanced Computing Center. Currently, he works on the PerfExpert project at the University of Texas at Austin. Prior to joining UT Austin, he held the position of a Performance Engineer at Salesforce.com.