SLIDE 1
Profiling
Program
Execution
Output Profile Data
- Feedback-directed optimization
- Code optimization
- ….
- Hardware optimization
- Hardware reconfiguration
- ….
Beyond Profiling Mohamed Zahran Chris Quackenbush Computer Science - - PowerPoint PPT Presentation
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
SLIDE 2
SLIDE 3
Hardware/Software Interaction
Start Execution End Execution
- No predictions
- Gather patterns
- Predict
- Update patterns
Training Inference
Examples: branch prediction, data value prediction, cache replacement, cache power management, …
SLIDE 4
Hypothesis
Software Hardware
Patterns
It is all about patterns. If we can learn the patterns, we can get useful info for ANY program and from ANY program.
SLIDE 5
Implementation
Offline Online
PARSEC
- List traversal
- Matrix transpose
- Array multiplication
- Quicksort
SLIDE 6
SLIDE 7
SLIDE 8
SLIDE 9
Roadmap
More sensitivity analysis and signature formats More test cases Branch prediction with offline data (Spectre?) Automatically generating patterns Continuous learning and re-learning
SLIDE 10
Conclusions
- The hardware (re)configuration depends
- n patterns.
- We can extra these patterns from any
profile data.
- The more profile data we have the more
patterns we learn.
SLIDE 11