ExpressPass: End-to-End Credit-based Congestion Control for Datacenters

TL;DR

ExpressPass is a novel end-to-end credit-based congestion control algorithm for datacenter networks that achieves rapid convergence, zero data loss, and low buffer occupancy, outperforming existing algorithms especially at high link speeds.

Contribution

It introduces a new credit-based congestion control mechanism that operates end-to-end, enabling faster convergence and better performance than prior algorithms like DCTCP.

Findings

Converges up to 80 times faster than DCTCP at 10Gbps links.

Reduces flow completion time for small and medium flows.

Achieves zero data loss and low buffer occupancy.

Abstract

As link speeds increase in datacenter networks, existing congestion control algorithms become less effective in providing fast convergence. TCP-based algorithms that probe for bandwidth take a long time to reach the fair-share and lead to long flow completion times. An ideal congestion control algorithms for datacenter must provide 1) zero data loss, 2) fast convergence, and 3) low buffer occupancy. However, these requirements present conflicting goals. For fast convergence,flows must ramp up quickly, but this risks packet losses and large queues. Thus, even the state-of-the-art algorithms, such as TIMELY and DCQCN, rely on link layer flow control (e.g.,Priority-based Flow Control) to achieve zero loss. This paper presents a new approach, called ExpressPass, an end-to-end credit-based congestion control algorithm for datacenters. ExpressPass is inspired by credit-based flow control, but extends it to work end-to-end. The switches control the amount of credit packets by rate limiting and ensure data packets flow in the reverse direction without any loss. ExpressPass leverages this to ramp up aggressively. ExpressPass converges up to 80 times faster than DCTCP at 10Gbps link, and the gap increases as link speeds become faster. Our simulation with realistic workload shows that ExpressPass significantly reduces the flow completion time especially for small and medium size flows compared to DCTCP, HULL, and DX.