Beyond Profiling Mohamed Zahran Chris Quackenbush Computer Science - PowerPoint PPT Presentation

Beyond Profiling Mohamed Zahran Chris Quackenbush Computer Science Dept. Google NYU cquackenbush@gmail.com mzahran@cs.nyu.edu

Profiling Output Execution Program Profile Data • Hardware optimization • Hardware reconfiguration • …. • Feedback-directed optimization • Code optimization • ….

Hardware/Software Interaction Start End Execution Execution • • No predictions Predict • • Gather patterns Update patterns Inference Training Examples: branch prediction, data value prediction, cache replacement, cache power management, …

Hypothesis Patterns Software Hardware It is all about patterns. If we can learn the patterns, we can get useful info for ANY program and from ANY program.

Implementation • List traversal Offline • Matrix transpose • Array multiplication • Quicksort Online PARSEC

Roadmap Continuous learning and re-learning Automatically generating patterns Branch prediction with offline data (Spectre?) More test cases More sensitivity analysis and signature formats

Conclusions • The hardware (re)configuration depends on patterns. • We can extra these patterns from any profile data. • The more profile data we have the more patterns we learn.

Questions! Chris Quackenbush Mohamed Zahran Google mzahran@cs.nyu.edu cquackenbush@gmail.com http://www.mzahran.com