Overview
Explore the intricacies of measuring and optimizing tail latency in data centers through this 52-minute conference talk from Strange Loop. Dive into the challenges of engineering interactive user request systems to optimize 99th percentile response times. Learn about a new tool and methodology for measuring performance at 1000 cycle granularities with minimal overhead. Discover root causes of tail latency and various optimization techniques, including scaling approaches for queuing delay and a novel dynamic adaptive parallelization method. Understand how these optimizations can improve server efficiency, benefiting users, profitability, and the environment. Gain insights from Kathryn McKinley, a Research Scientist at Google and Adjunct Professor at the University of Texas at Austin, as she shares her expertise in programming languages, compilers, runtime systems, and performance optimization.
Syllabus
Introduction
We are impatient people
Tail latency
Big data centers
Energy consumption
Cost savings
Efficiency
Server Architecture
Overview
Long Requests
Simplified Requests
Prior State of the Art
Memory Locations
Shared Locations
Log of Samples
Configuration
Program Counter
Shortening Queue
Parallelization
Slow to Fast
Pegasus
Results
Conclusion
Questions
Taught by
Strange Loop Conference