Backpressure flow control

Abstract

Effective congestion control in a multi-tenant data center is becoming increasingly challenging with rapidly increasing workload demand, ever faster links, small average transfer sizes, extremely bursty traffic, limited switch buffer capacity, and one-way protocols such as RDMA. Existing deployed algorithms, such as DCQCN, are still far from optimal in many plausible scenarios, particularly for tail latency. Many operators compensate by running their networks at low average utilization, dramatically increasing costs. In this paper, we argue that we have reached the practical limits of end-to-end congestion control. Instead, we propose a new clean slate design based on hop-by-hop per-flow flow control. We show that our approach achieves near optimal tail latency behavior even under challenging conditions such as high average link utilization and in-cast cross traffic. By contrast with prior hop-by-hop schemes, our main innovation is to show that per-flow flow control can be achieved with limited metadata and packet buffering. Further, we show that our approach generalizes well to cross-data center communication.