HPCC

Abstract

Congestion control (CC) is the key to achieving ultra-low latency, high bandwidth and network stability in high-speed networks. From years of experience operating large-scale and high-speed RDMA networks, we find the existing high-speed CC schemes have inherent limitations for reaching these goals. In this paper, we present HPCC (High Precision Congestion Control), a new high-speed CC mechanism which achieves the three goals simultaneously. HPCC leverages in-network telemetry (INT) to obtain precise link load information and controls traffic precisely. By addressing challenges such as delayed INT information during congestion and overreaction to INT information, HPCC can quickly converge to utilize free bandwidth while avoiding congestion, and can maintain near-zero in-network queues for ultra-low latency. HPCC is also fair and easy to deploy in hardware. We implement HPCC with commodity programmable NICs and switches. In our evaluation, compared to DCQCN and TIMELY, HPCC shortens flow completion times by up to 95%, causing little congestion even under large-scale incasts. CCS CONCEPTS• Networks → Transport protocols; Data center networks;

SDCCP: Control the network using software‐defined networking and end‐to‐end congestion control

“…Hybrid schemes leverage In‐Network Technology to notify end‐to‐end hosts the network condition by using ECN, 37,38 piggybacking on ACK packets, 39 or adding additional packet fields 32 …”

Congestion Control for Datacenter Networks: A Control-Theoretic Approach

“…Typical datacenter networks have topologies with fixed distances (in contrast to the Internet) and fixed bisection bandwidth (in contrast to wireless networks), which may make congestion control there seem simple. It turns out to be quite the opposite, though, as evidenced by the spectrum of solutions that exploit different congestion signals [1], [2], leverage latest developments in network hardware [3], and revisit previous work with a new perspective [4].…”

“…RDMA requires losslessness, triggering the need for PFC (priority-based flow control) [10], which prevents packet drop by using back-pressure (at the traffic class level). Alas, PFC has been observed to cause problems such as HoL (head-of-line) blocking, congestion spreading, and routing deadlocks [1], [2], [4], [3], [11]. Less aggressive flow control mechanisms [12] have been proposed to replace PFC.…”

“…However, we believe that PFC should only be triggered to prevent buffer overrun, and we show that if congestion control is able to maintain stable queues on switches, then PFC activation is rare. Datacenter networks have failed to harness the full potential of RDMA due to inefficient congestion control [3], and the many RDMA congestion control solutions developed over the recent past, e.g., [1], [2], [4], [3], are indicators that congestion control for RDMA is a critical problem.…”

“…The recently proposed HPCC (high precision congestion control) [3] is a source-driven, window-based solution leveraging INT (in-band network telemetry) to gather link load information and adjust source-side transmission window sizes. HPCC outperforms DCQCN, but fails to meet fairness guarantees in scenarios -as we show ( § VI-A) -commonly observed in modern datacenter networks [1].…”