How DAP Beats Traditional Congestion Control in Real-Time Video Streaming

At the ACM Multimedia 2021 conference in Chengdu, the Kuaishou audio‑video network transmission algorithm team (Kwai2021) won the ACM MM Grand Challenge: Meet Deadline Requirements with their DAP system, and the related paper was accepted to the conference.

Background

Real‑time communication (RTC) applications such as video conferencing and online collaboration have become ubiquitous, especially after the pandemic, demanding ultra‑low end‑to‑end latency. RTC traffic consists of three data types—signaling, video, and audio—each with different priority and latency tolerance.

DAP System Overview

DAP comprises a sender and a receiver. The sender processes network statistics and dispatches data using two core modules: a scheduling module that decides which data block to transmit, and a congestion‑control module that adjusts the sending rate. The receiver gathers media packets, measures metrics such as loss rate and receive speed, and feeds them back to the sender.

The scheduling algorithm selects data blocks based on a reward function that considers block importance, deadline, packet loss rate, and available bandwidth, quantifying each block’s contribution to user experience. Only packets sent before their deadline count toward QoE.

The congestion‑control algorithm adopts an optimized “paired‑sending” strategy to estimate packet transmission time and dynamically adjust the number of packets sent simultaneously according to network reordering. This approach equalizes packet send times, mitigates clock drift, and improves the accuracy of delay measurements.

Additionally, DAP may intentionally “over‑send” packets when the network bandwidth is underestimated, accelerating bandwidth probing and increasing throughput while keeping latency low.

Performance Evaluation

On the official ACM MM Grand Challenge datasets, DAP outperformed state‑of‑the‑art congestion‑control algorithms. Compared with Copa, BBR, and a deep‑reinforcement‑learning (DRL) based method, DAP improved user experience by 43.5%, 39.12%, and 30.31% respectively, achieving a final score of 5012.26 and securing first place.

Ablation studies confirmed that both the deadline‑aware scheduling and the paired‑sending congestion control contribute to DAP’s superiority. When combined with existing scheduling strategies (Deadline‑First, Priority‑First), DAP’s scheduling still yields significant gains, and its congestion‑control component consistently outperforms other algorithms in both throughput and latency.

Team

The winning team is Kuaishou’s audio‑video technology department, responsible for optimizing network and transmission performance across short video, live streaming, and RTC scenarios. Their work spans transmission protocols, congestion control, weak‑network mitigation, ABR, and reinforcement learning, with multiple publications in top venues such as INFOCOM, ACM MM, TON, TMC, and JSAC.