A Distributed Algorithm to Calculate Max-Min Fair Rates Without Per-Flow State

Abstract

Most congestion control algorithms, like TCP, rely on a reactive control system that detects congestion, then marches carefully towards a desired operating point (e.g. by modifying the window size or adjusting a rate). In an effort to balance stability and convergence speed, they often take hundreds of RTTs to converge; an increasing problem as networks get faster, with less time to react. This paper is about an alternative class of congestion control algorithms based on proactive-scheduling: switches and NICs "pro-actively" exchange control messages to run a \em distributed algorithm to pick "explicit rates for each flow. We call these Proactive Explicit Rate Control (PERC) algorithms. They take as input the routing matrix and link speeds, but not a congestion signal. By exploiting information such as the number of flows at a link, they can converge an order of magnitude faster than reactive algorithms. Our main contributions are (1) s-PERC ("stateless" PERC), a new practical distributed PERC algorithm without per-flow state at the switches, and (2) a proof that s-PERC computes exact max-min fair rates in a known bounded time, the first such algorithm to do so without per-flow state. To analyze s-PERC, we introduce a parallel variant of standard waterfilling, 2-Waterfilling. We prove that s-PERC converges to max-min fair in 6N rounds, where N is the number of iterations 2-Waterfilling takes for the same routing matrix. We describe how to make s-PERC practical and robust to deploy in real networks. We confirm using realistic simulations and an FPGA hardware testbed that s-PERC converges 10-100x faster than reactive algorithms like TCP, DCTCP and RCP in data-center networks and 1.3--6x faster in wide-area networks (WANs). Long flows complete in close to the ideal time, while short-lived flows are prioritized, making it appropriate for data-centers and WANs.