ACC: agile congestion control in datacenter networks

Faced with the increasingly complex and rapidly changing datacenter network traffic, congestion control technology based on INT (in-network telemetry) information faces many challenges. On the one hand, the existing state-of-the-art algorithms mainly rely on the receiver's ACK (acknowledge) packet to feedback INT information to the sender to obtain link load information. Since the feedback delay of the ACK packet is at the RTT (round trip time) level, the sender cannot obtain INT information within the first RTT of the flow transmission. On the other hand, the existing algorithms generally lack the ability to seize the idle network bandwidth when the flow transmission is completed, so the network would inevitably waste bandwidth. In this paper, we explore INT-based techniques, break the one-RTT barrier efficiently and eliminate bandwidth waste. We come up with an ultra-precise and efficient congestion control algorithm, called ACC (agile congestion control). Compared with the existing state-of-the-art INT-based congestion control algorithm HPCC (high precision congestion control), ACC pays special attention to the first RTT of the flow transmission. During this period, the switch is used to transmit the path congestion signal back to the sender, which can respond more quickly to network congestion. In addition, ACC pays attention to the bandwidth preemption ability of long flows in the network, which can help long flows perceive idle bandwidth in advance, thereby reducing the tail delay of network transmission. Furthermore, we conduct extensive experiments to evaluate the performance of our design. The result shows that while keeping the same high throughput as HPCC, ACC can maintain a lower queue length and effectively reduce the tail delay of the flow. Specifically, in large-scale Clos topology, when the workload is WebServer and the workload intensity is 0.6, compared with HPCC, ACC can reduce the average queue length by 9.6%, shorten the 95th-percentile flow completion time by 29.7%.