Proactive Content Caching via Interplay Between Deep Learning and Stochastic Optimization
Yongmoon Park, Kyungtae Lee, Minseok Choi, Yeongjin Kim · 2024
The increasing demand for video data traffic, along with the proliferation of smart devices, poses significant chal-lenges to content and Internet service providers. In response to this challenge, content caching on mobile edge computing (MEC) servers has emerged to reduce content download latency. However, existing caching solutions often assume stationary content popularity or require real-time knowledge of popularity, which does not align with real-world scenarios. To address these limitations, we introduce Pro Cache, a novel content caching algorithm. ProCache takes into account spatial and temporal monetary budget sharing for caching, content sizes, and original server locations while dealing with uncertain regional content popularity. The goal of ProCache is to minimize the long-term expected content download latency overall active users, comprising two key components. First, the deep learning module includes the IDCNN-LSTM-Dense layered deep learning model for predicting future content requests and the data mapping module which makes the model focus more on the request trend rather than the magnitude. Second, based on our predictions, the stochastic optimization module runs a dynamic content caching algorithm based on the Lyapunov optimization technique that operates in a fully distributed manner by region and ensures overall performance bounds. Trace-driven simulations using the YouTube dataset demonstrate that ProCache outperforms existing prediction models and content caching algorithms.