Performance analysis and optimization for extreme scale systems
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Thanks to the death of Dennard scaling and the slowing of Moore’s Law, future compute platforms are becoming increasingly complex as they attempt to continue to scale either overall performance or performance efficiency. Consequently, analyzing the performance of such systems has also become an ever-more complex problem. Gone
are the days where peak FLOPS numbers were a reliable indicator of real-world performance, and they have been replaced by a field where performance bottlenecks can occur in any component of these platforms. The reality of the situation requires that performance analysis be conducted at many levels throughout the lifecycle of application development, focusing on both individual elements of computation and also end-to-end system performance of the application at scale, as well as many additional points within that spectrum. In this work, we present both component-level and system-level performance analysis tools, as well as a non-traditional application of using online performance analysis to enable dynamic performance optimization.
Description
Keywords
Performance analysis, Optimization, Extreme scale systems