Overview
Explore a groundbreaking approach to estimating tail latency performance in large-scale data center networks through this conference talk from NSDI '23. Discover how researchers from the University of Washington, Microsoft Research, and MIT CSAIL tackle the challenge of providing fast and accurate estimates for flow-level tail latency, a crucial metric for cloud application performance. Learn about the innovative techniques developed to decompose the problem into parallel independent single-link simulations, enabling accurate estimates of end-to-end flow level performance distributions for entire networks. Understand how this method achieves significant speedups compared to traditional network simulators like ns-3 and OMNeT++, running in minutes rather than hours or days. Gain insights into the approach's ability to handle general traffic matrices and topologies without relying on machine learning, eliminating training delays while maintaining accuracy within 9% for tail flow completion times.
Syllabus
NSDI '23 - Scalable Tail Latency Estimation for Data Center Networks
Taught by
USENIX