Explore automated performance tuning techniques using Bayesian optimization in this conference talk from Twitter engineers Joshua Cohen and Ramki Ramakrishna. Dive into the challenges of managing resource utilization in Twitter's Mesos clusters and learn how machine learning can efficiently search large parameter spaces to optimize specific performance metrics. Discover the development of a system for continuous automated tuning of services, addressing the complexities of multitudinous knobs, heterogeneous hardware, and diverse service requirements. Gain insights into the JVM layer tuning, microservice optimization, and the integration of Apache Aurora for experiment management. Understand the setup, evaluation, and results of performance optimization efforts, including improvements in GC cost and overall service efficiency. Examine the concept of AutoTune as a service and its potential applications beyond the JVM, concluding with future directions for this innovative approach to resource management and performance enhancement in large-scale distributed systems.
Automated Performance Tuning with Bayesian Optimization
Overview
Syllabus
Intro
TWITTER RUNS ON MICROSERVICES
A PERFORMANCE STACK AT TWITTER
TUNING AT THE JVM LAYER
PERFORMANCE OPTIMIZATION
CONSTRAINTS
PERFORMANCE TUNING
OPTIMIZATION OF A BLACK BOX FUNCTION
BAYESIAN OPTIMIZATION EXAMPLE
ALTERNATIVE APPROACHES
BAYESIAN OPTIMIZATION EXPERIENCES AT TWITTER
MICROSERVICE STACK
OPTIMIZING A MICROSERVICE BY TUNING THE JVM
A SAMPLING OF JVM PARAMETERS
SET-UP
EVALUATION
PERFORMANCE OF THE OPTIMUM RESULT
GC COST
OPTIMIZED SETTINGS
KEY TAKEAWAYS
AUTOTUNE AS A SERVICE
WHAT DOES AURORA BRING TO THE TABLE
AURORA BASICS
LAUNCHING AN EXPERIMENT
A BRIEF DIVERSION
RUNNING AN EXPERIMENT
FINISHING AN EXPERIMENT
CLOSING THE LOOP
THE VIRTUOUS CIRCLE
BEYOND THE JVM
CONCLUSION
WHAT'S NEXT
Taught by
Linux Foundation
