Designing a congestion control plane datapath with QUIC

Abstract

This work explores developing datapaths for the recently proposed Congestion Control Plane (CCP) architecture and uses QUIC as a case study to build and evaluate an example datapath. The CCP moves congestion control logic off of the datapath into a separate agent running in user-space. Now, algorithm developers can write their algorithm using this API and automatically run their algorithm on any CCP enabled datapath, such as QUIC or the Linux kernel. We discuss the necessary features for datapaths to support the user-space CCP agent and develop a library, libccp, for software datapaths. We use QUIC as a case study to design a CCP datapath and evaluate various congestion control algorithms running on top of QUIC. The evaluation focuses on four aspects: (1) Do algorithms written on CCP have the same performance as QUIC algorithms? (2) Do CCP congestion control algorithms behave similarly to their native QUIC counterparts, when run on top of QUIC? (3) How expressive is the CCP API? (4) Does the same algorithm, written through CCP, behave similarly across multiple datapaths? This work reveals that while CCP QUIC algorithms have similar performance for a single flow, their behavior does not exactly match the native QUIC implementations of the same algorithms without modifications. We show that CCP can provide "write once, run anywhere" semantics for congestion control, as multiple algorithms run across QUIC and the Linux kernel exhibit similar behavior. Finally, we evaluate the expressiveness of the CCP API by implementing two algorithms: Hybrid Slow Start and Remy.