PRED - Performance-oriented Random Early Detection for Consistently Stable Performance in Datacenters

Learn about PRED, a novel system for automatic and stable Random Early Detection (RED) parameter adjustment in datacenter networks through this 16-minute conference presentation from NSDI '25. Discover how researchers from Huawei Technologies, Renmin University of China, and Tsinghua University address the challenge of maintaining consistently high performance with dynamically configured RED thresholds in highly dynamic datacenter workloads. Explore the two loosely coupled systems - Flow Concurrent Stabilizer (FCS) and Queue Length Adjuster (QLA) - that enable PRED to overcome the limitations of static RED configurations and learning-based methods that suffer from poor tail performance due to instability. Examine the extensive evaluation results from physical testbed experiments and large-scale simulations demonstrating PRED's ability to keep up with real-time network dynamics, achieving 66% lower switch queue length and up to 80% lower Flow Completion Time compared to static-threshold-based methods, while reducing tail FCT by 34% compared to state-of-the-art learning-based approaches.