Low-Latency Video Conferencing via Optimized Packet Routing and Reordering

Yao Xiao, Sitian Chen, Amelie Chi Zhou, Shuhao Zhang, Yi Wang, Rui Mao, Xuan Yang

TL;DR

This paper tackles the challenge of low-latency video conferencing across geographically distributed data centers by introducing VCRoute, a novel routing algorithm that jointly optimizes for packet transmission latency and its variance, and WMJitter, a watermark-based mechanism for efficient jitter management. These techniques significantly reduce end-to-end latency, improving the quality of experience for users in global video calls.

What Problem Does This Paper Address?

The explosion of video conferencing, particularly since the COVID-19 pandemic, has highlighted a critical challenge: maintaining low latency and high quality of service across geographically distributed data centers (geo-distributed DCs). Unlike local networks, the connections between DCs (inter-DC networks) are characterized by high latency (hundreds of milliseconds) and significant dynamism. This means that network conditions can change rapidly, leading to unpredictable delays (jitter) and packets arriving out of order. Existing solutions often focus on minimizing transmission latency alone, which doesn't guarantee low end-to-end latency because it doesn't account for the time packets spend being reordered at the receiver. Furthermore, traditional jitter buffers, while effective, can introduce their own delays by waiting for all packets to arrive in order. This paper aims to bridge this gap by optimizing both the routing of packets and their subsequent reordering to achieve truly low-latency video conferencing.

Methods

The paper introduces two core innovations:

  • VCRoute (Video Conferencing Routing): This is an application-aware routing method designed for audio data packets. It frames the routing problem as a Multi-Armed Bandit (MAB) problem, where each potential path between endpoints is an "arm." Crucially, VCRoute adapts the Thompson Sampling algorithm to this MAB framework. Unlike traditional MAB approaches that might only consider average latency, Thompson Sampling allows VCRoute to model the distribution of latency for each path, enabling it to simultaneously account for both the average transmission latency and its variance. This is vital for dynamic inter-DC networks where high variance can negate the benefits of low average latency. To manage overhead, VCRoute first identifies a set of "top-k" promising paths based on historical data and then dynamically selects the best path for each packet from this reduced set.
  • WMJitter (Watermark-based Jitter Manager): This is a novel jitter management mechanism implemented at the receiver's end. It moves away from traditional In-Order Processing (IOP) jitter buffers towards an Out-Of-Order Processing (OOP) paradigm using watermarks. Watermarks are used to indicate the progress of the data stream. Packets arriving before the watermark are considered timely, while those arriving significantly later are discarded. This watermark-based approach allows for more flexible and efficient handling of out-of-order packets, reducing the reordering latency and overall delay compared to conventional methods that wait for all packets to arrive in sequence. WMJitter also dynamically adjusts a "lag" parameter to balance latency and packet loss, similar to how buffer sizes are adjusted in traditional systems.

Datasets

The evaluation of VCRoute and WMJitter was performed using network performance traces collected from two real-world geo-distributed environments:

  • Tencent Cloud: This platform provided traces from five data centers, with users located in Shenzhen, Hong Kong, and Singapore.
  • WonderProxy: This platform offered a global network of 250 servers, with traces collected from five servers acting as endpoints and five as users across various global locations (e.g., Brazil, Australia, Singapore, South Africa, Greece).

The authors did not introduce new benchmarks but rather used these existing trace datasets to simulate realistic geo-distributed network conditions.

Key Insights and Ideas

  • Latency is More Than Just Transmission Time: The paper strongly emphasizes that for real-time applications like video conferencing, minimizing packet transmission latency alone is insufficient. The time spent reordering packets due to jitter is a significant contributor to end-to-end latency, and this must be addressed.
  • Variance Matters in Dynamic Networks: In highly dynamic inter-DC networks, focusing solely on average latency can be misleading. VCRoute's use of Thompson Sampling to consider latency variance is a key insight for making robust routing decisions.
  • Out-of-Order Processing for Lower Latency: The WMJitter approach demonstrates that by embracing out-of-order processing, it's possible to reduce jitter-induced delays without sacrificing ordered delivery, a significant departure from traditional jitter buffers.
  • Joint Optimization is Key: The paper's success stems from the synergistic combination of VCRoute (optimizing the path) and WMJitter (optimizing packet handling at the receiver), showing that a holistic approach is necessary for optimal performance.

Key Results

The proposed VCRoute and WMJitter system demonstrated significant improvements:

  • End-to-End Latency Reduction: VCR-WM (VCRoute + WMJitter) achieved up to 44% reduction in end-to-end packet latency compared to state-of-the-art methods.
  • Routing Performance: VCRoute outperformed existing routing methods like "Via" by 4%-13% in reducing average end-to-end latency, especially in heterogeneous network conditions.
  • Jitter Management Effectiveness: WMJitter provided substantial latency improvements (e.g., 8%-16% reduction) over traditional buffer-based methods, particularly in highly dynamic network environments.
  • Loss Rate: VCR-WM yielded lower or comparable packet loss rates while delivering significantly lower latency.
  • System Overhead: The additional overhead introduced by VCRoute was less than 1% of the average end-to-end latency, indicating good scalability.

Limitations and Future Work

  • Centralized VCRoute: While VCRoute is designed to be centralized for efficiency, the paper acknowledges the potential overhead of transmitting routing decisions. Future work could explore decentralized or hierarchical approaches to further mitigate this.
  • Application Specificity: The current VCRoute is designed for audio data packets. Extending it to handle video packets, which have different characteristics and latency sensitivities, would be a valuable next step.
  • Real-time Deployment: The evaluation was conducted using trace-driven simulations. Real-world deployment and testing in live video conferencing systems would provide further validation.
  • Network Dynamics: While the methods are designed for dynamic networks, exploring more advanced adaptive strategies for even more extreme or rapidly changing network conditions could be beneficial.

Summary updated

Abstract

In the face of rising global demand for video meetings, managing traffic across geographically distributed (geo-distributed) data centers presents a significant challenge due to the dynamic and limited nature of inter-DC network performance. Facing these issues, this paper introduces two novel techniques, VCRoute and WMJitter, to optimize the performance of geo-distributed video conferencing systems. VCRoute is a routing method designed for audio data packets of video conferences. It treats the routing problem as a Multi-Armed Bandit issue, and utilizes a tailored Thompson Sampling algorithm for resolution. Unlike traditional approaches, VCRoute considers transmitting latency and its variance simultaneously by using Thompson Sampling algorithm, which leads to effective end-to-end latency optimization. In conjunction with VCRoute, we present WMJitter, a watermark-based mechanism for managing network jitter, which can further reduce the end-to-end delay and keep an improved balance between latency and loss rate. Evaluations based on real geo-distributed network performance demonstrate the effectiveness and scalability of VCRoute and WMJitter, offering robust solutions for optimizing video conferencing systems in geo-distributed settings.
BibTeX
@inproceedings{xiao2024low,
  author = {Yao Xiao and Sitian Chen and Amelie Chi Zhou and Shuhao Zhang and Yi Wang and Rui Mao and Xuan Yang},
  title = {{Low-Latency Video Conferencing via Optimized Packet Routing and Reordering}},
  booktitle = {IWQoS},
  year = {2024},
  url = {https://openreview.net/forum?id=0GrvgBh3hu},
}
Figure