Learning Driven Elastic Task Multi-Connectivity Immersive Computing Systems

TL;DR

This paper proposes a learning-based framework for elastic task offloading in multi-user VR systems with multi-connectivity, optimizing energy efficiency and reducing latency using centralized and decentralized methods trained on real network data.

Contribution

It introduces a novel centralized and decentralized reinforcement learning framework for elastic VR task offloading, addressing energy efficiency, latency, and system scalability challenges.

Findings

CPPG reduces latency by 28%.

CPPG decreases energy consumption by 78%.

The methods are effective on real-world network traces.

Abstract

In virtual reality (VR) environments, computational tasks exhibit an elastic nature, meaning they can dynamically adjust based on various user and system constraints. This elasticity is essential for maintaining immersive experiences; however, it also introduces challenges for communication and computing in VR systems. In this paper, we investigate elastic task offloading for multi-user edge-computing-enabled VR systems with multi-connectivity, aiming to maximize the computational energy-efficiency (computational throughput per unit of energy consumed). To balance the induced communication, computation, energy consumption, and quality of experience trade-offs due to the elasticity of VR tasks, we formulate a constrained stochastic computational energy-efficiency optimization problem that integrates the multi-connectivity/multi-user action space and the elastic nature of VR computational tasks. We formulate a centralized phasic policy gradient (CPPG) framework to solve the problem of interest online, using only prior elastic task offloading statistics (energy consumption, response time, and transmission time), and task information (i.e., task size and computational intensity), while observing the induced system performance (energy consumption and latency). We further extend our approach to decentralized learning by formulating an independent phasic policy gradient (IPPG) method and a decentralized shared multi-armed bandit (DSMAB) method. We train our methods with real-world 4G, 5G, and WiGig network traces and 360 video datasets to evaluate their performance in terms of response time, energy efficiency, scalability, and delivered quality of experience. We also provide a comprehensive analysis of task size and its effect on offloading policy and system performance. In particular, we show that CPPG reduces latency by 28% and energy consumption by 78% compared to IPPG.