3.3.2 Tunable Optimization Objective Engine

[Mathnerd314]

Desired function
Provide a scalar cost for candidate codegen/optimization decisions, incorporating performance and resource objectives.

Futamura role
Used by compiler (and by online JIT decisions) to select among multiple valid compiled results.

Inputs

• User weights per metric / overall weighting function of performance and resources objectives
• Piece of code, profiling data

Processing

• Record offline statistics of program performance - large corpus, should include:
  • run time (wall time, CPU time, etc.)
  • code size
  • compile time
  • memory usage
  • power
  • latency (of main loop, like server requests per second)
• Build ML model and fit statistical distributions and their parameters e.g. mean/variance/tail - should try to get good fits
• Using modeling to normalize metrics to common/standardized scale (e.g. percentiles)
• Combine via weighted linear combination or lexicographic ordering.
• Optionally include risk terms (measures of variance / tail spread of predicted distribution, use non-median measures of performance like 10th percentile)

Outputs

• Provide scalar cost for search/heuristics.
• Provide simple sensitivity analysis / breakdown - contributions of each metric to final score (observability)
• drives copy-and-patch vs. advanced tracing/loop analysis decisions, other optimization decisions