On Modular Learning of Distributed Systems for Predicting End-to-End Latency

Explore a cutting-edge approach to modeling end-to-end latency in distributed systems through this 19-minute conference talk from NSDI '23. Discover Fluxion, a novel framework that employs modularized learning to characterize system performance more efficiently and accurately than traditional monolithic models. Learn how this innovative method significantly reduces adaptation costs in dynamic cloud deployments, allowing for faster model updates and improved system performance tuning. Gain insights into the implementation of learning assignments, a new abstraction that enables modeling of individual sub-components with a consistent interface. Examine the framework's effectiveness through case studies involving complex systems with up to 142 microservices, demonstrating substantial improvements in prediction accuracy and system optimization.

NSDI '23 - On Modular Learning of Distributed Systems for Predicting End-to-End Latency