Cross-layer host-network co-design for QoS in streamed video data networking
With the surge in video applications, video traffic has become the dominant form of Internet traffic. Given their significant data size and the need for low-latency transmission, video data present considerable challenges to contemporary networking systems. We identify a pivotal research gap as the...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis-Doctor of Philosophy |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/178407 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | With the surge in video applications, video traffic has become the dominant form of Internet traffic. Given their significant data size and the need for low-latency transmission, video data present considerable challenges to contemporary networking systems. We identify a pivotal research gap as the disjoint design of hosts and the network. To address this, we introduce the concept of host-network co-design, aiming at the enhancement of system performance through the coordination between hosts and the network. The research undertaken during my PhD delineates host-network co-design in two distinct scenarios, namely video analytics systems and general-purpose wide area networks.
Our first work addresses the problem of excessive bandwidth consumption for data transmission and computation constraints on edge servers of existing video analytics systems. A cloud-edge collaborative architecture is proposed to combine edge-based inference with cloud-assisted continuous learning. Lightweight DNN models are maintained at the edge servers and continuously retrained with a more comprehensive model on the cloud to achieve high video analytics performance while reducing the amount of data transmitted between edge servers and the cloud. The proposed design faces the challenge of constraints of both computation resources at the edge servers and network bandwidth of the edge-cloud links. An accuracy gradient-based resource allocation algorithm is proposed to allocate the limited computation and network resources across different video streams to achieve the maximum overall performance. A prototype system is implemented and experiment results demonstrate the effectiveness of our system with up to 28.6% absolute mAP gain compared with alternative designs.
Our second work takes the perspective of a general-purpose Internet service provider, where various host applications have uncertain latency objectives which vary both across applications and over time. The traditional WAN operates in a way isolated from hosts, which fails to address such uncertainty. We propose PredWAN, a host-network synergistic WAN which enables application-specified service differentiation. We extend the QUIC protocol to support QUIC CID coding, which encodes per-packet latency objectives in QUIC CIDs, to allow host applications to convey the real-time requirements of their transmission latency to the WAN. The WAN can schedule packets to routing paths with corresponding latencies to satisfy such objectives. The transmission path and latency are enforced while maintaining transparency to application servers through SRv6 tunneling. Moreover, a hardware offloading solution based on hash compression is applied to efficiently accelerate packet processing and reduce computation overhead introduced by the system, which is compatible to both the QUIC standard and existing hardware specifics. PredWAN is evaluated with thorough trace-driven case studies. Results indicate that PredWAN outperforms the traditional WAN under various network conditions and application scenarios, and is practical for real-world deployment.
Both works share the same thought, which is to use application-specified knowledge available on hosts to guide the data transmission in the network for latency-sensitive data such as video streaming. In this way, the network no longer transmits data indiscriminately but is utilized in an application-aware way. We consider the host-network synergy approach provides a transformative shift in networking paradigms that improves both performance and efficiency. |
---|