Back to index

Finding latent performance bugs in systems implementations

Charles Killian, Karthik Nagaraj, Salman Pervez, Ryan Braud, James W. Anderson, and Ranjit Jhala
Purdue University and UCSD

One-line Summary

MACE Performance Checker is a model checking tool to detect latent performance bugs in the event-based distributed systems.

Overview/Main Points

Goal: detect latent performance bugs in corner cases of an atomic event-driven state machine model
High level idea: explore different timing to expose latent bugs

How to identify a latent bug?
What is provided by developers?
explore different timings: use a per-node clock ‘node, event, start, duration’

The algorithm
- Inputs
  - event queue: ‘node, event, start’
  - timedEvent queue: ‘node, event, start, duration’
  - realTimes queue: ‘event, realTime’
- Training
  - Event Duration Distributions (EDD) training
  - Overall performance training
- Abnormal detection
- Divergence detection
- Frequency correlation
Performance problems in distributed systems
- race-related latent bugs

Relevance

Flaws