Video Caching at Data-drifting Network Edge: A KD-based Cross-domain Collaborative Solution
Primary Area: infrastructure, software libraries, hardware, etc.
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Cooperative edge caching; Data drift; Knowledge distillation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: The explosive growth of video content streaming has led to network congestion and quality decline, making efficient content delivery a significant challenge. To address this, edge caching has emerged as a solution, utilizing mobile edge caching servers like edge base stations (EBS) as a cost-effective approach. Collaborative edge caching has been proposed to address the space limitation of edge servers by enabling cooperation and content sharing among multiple servers, thereby improving caching hit rates (CHR). However, little attention has been paid to the impact of request characteristics on different servers. To tackle this issue, we conducted a study using data collected from Kuaishou company over a period of four weeks, comprising 350 million video requests. Our research findings indicate that request-sparse EBSs significantly impede the overall CHR improvement in the edge caching problem. Knowledge distillation (KD), a technique that transfers knowledge from strong models to weak models, is expected to solve this bottleneck problem. However, traditional KD methods often rely on the assumption of independent and identically distributed data, which may not hold true in real-world scenarios where data drift occurs. We identify two major types of data drift in caching data: temporal drift and spatial drift. To overcome these challenges, we propose an adaptive KD-based cross-domain collaborative edge caching framework, called KDCdCEC, which incorporates three specifically tailored components: i) a slot-wise reinforcement learning agent capable of adapting to EBSs with varying storage sizes, ii) a deep deterministic policy gradient-based algorithm that adaptively configures the reference weights of servers on their partners, and iii) a content-aware request routing mechanism that enhances the decision-making of edge servers. Experimental results show that KDCdCEC outperforms state-of-the-art approaches in terms of average CHR, average latency, and traffic cost.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7537
Loading