Performance-Detective: Automatic Deduction of Cheap and Accurate Performance Models

Explore an innovative approach to performance modeling for complex applications in this conference talk from the International Conference on Supercomputing (ICS 2022). Discover Performance-Detective, a code analysis tool that deduces insights on program parameter interactions to create efficient and accurate performance models. Learn how this method significantly reduces the number of measurements needed, cutting costs to just 2.9% of previously required core hours while maintaining model accuracy. Dive into case studies demonstrating the tool's effectiveness, and understand its potential impact on optimizing configuration choices for modern applications with numerous options. Follow the speaker's journey from motivation and current modeling workflows to the detailed workings of Performance-Detective, including system analysis, experiment design, and modeling techniques using decision trees.

Intro
Motivation: Cost of computing clusters
Motivation: Configuration options
Performance Modeling
Current Modeling Workflows
Performance-Detective Modeling Workflow
System analysis [Perf-Taint]
Pace3D: System analysis
Pace3D: Experiment Design
RQ1: Modeling using a single repetition of experiments
simultaneously - Modeling with Decision Trees