What You See Is Not What You (Would) Get: Adventures in Parallel Performance Overhead Compensation
|Author:||Allen D. Malony University of Oregon|
|Date:||October 27, 2005|
Parallel performance profiling and tracing generates measurement overhead during parallel program execution. Measurement overhead, in turn, introduces intrusion in a program's runtime performance behavior. Intrusion can be mitigated by controlling instrumentation degree, allowing a tradeoff of accuracy for detail. Alternatively, the accuracy in performance results can be improved by reducing the intrusion error due to measurement overhead. Models for compensation of measurement overhead in parallel performance profiling and tracing are described. An approach based on rational reconstruction is used to understand properties of compensation solutions for different parallel scenarios. From this analysis, a general algorithm for on-the-fly overhead assessment and compensation is derived for parallel profiling. For tracing, the algorithms are run post-mortem, allowing for the use of additional performance information (modeled or derived) to improve parallel performance estimates.
The techniques for compensation analysis in parallel performance profiling and tracing have been implemented in the TAU parallel performance system and the EPILOG/Expert trace analysis package. This is described and we highlight the implementation problems for MPI message passing programs. Experimental results are shown for the NAS parallel benchmarks and other applications demonstrating that overhead compensation can be effective in improving the accuracy of measured performance data. For one case study, it is shown that the error due to measurement intrusion can be reduced from 85% (what you see) to less than 1% (what you get).
Dr. Allen D. Malony is a Professor in the Department of Computer and Information Science at the Unversity of Oregon. He received the Ph.D. degree from the University of Illinois at Urbana-Champaign in October 1990, joining the faculty at Oregon in 1991. Dr. Malony was a Fulbright Research Scholar to The Netherlands in 1991 and to Austria in 1999. He was awarded the NSF National Young Investigator award in 1994 and the Alexander von Humboldt Research Award for Senior U.S. Scientists by the Alexander von Humboldt Foundation in 2002.
Dr. Malony's research interests are in parallel computing, performance analysis, supercomputing, and scientific software environments. He is the Director of the Performance Research Laboratory at the University of Oregon where he manages research and development projects funded by the Department of Energy and the National Science Foundation. Dr. Malony's research interests extend to computing and information processing systems for neuroscience research. He is the Director of the NeuroInformatics Center (NIC) at the University of Oregon. The NIC is developing advanced integrated neuroimaging tools that combine EEG and MRI methods for next-generation brain analysis. Grid technologies and high-performance computing are being used by the NIC to prototype network-based systems for medical service delivery.