Resource Management for Edge-Assisted Extended Reality

Abstract

Extended reality (XR) enables immersive experiences by seamlessly merging the physical and digital worlds. Supporting such experiences requires real-time and high-quality rendering of virtual content to generate video frames, which is computationally intensive and poses a challenge for resource-constrained XR devices. To overcome this limitation, a promising approach is to offload rendering tasks to nearby edge servers with powerful computing resources. In an edge-assisted XR system, interdependent tasks, including video frame rendering, encoding, and transmission, need to be executed in a pipeline, which consumes substantial communication, computing, and caching resources. The efficiency of network resource provisioning has a direct impact on users' quality of experience (QoE), which reflects the presence and immersive of a user during virtual content viewing and is measured by the weighted sum of visual quality, quality variation, and round-trip latency. Our objective is to efficiently manage multi-dimensional network resources for the XR service to improve user QoE under dynamic network environments. However, the technical challenges are as follows: 1) given the spatiotemporally varying service demand caused by user mobility, how to proactively provision edge resources for the service while achieving satisfactory user QoE; 2) how to adaptively allocate edge resources for individual users to accommodate demand fluctuations caused by dynamic viewing behavior; and 3) in the presence of task dependencies in the pipeline, how to jointly coordinate task processing parameters (e.g., rendering quality, frame encoding type) to improve user QoE.

In this thesis, we design efficient resource management schemes for an edge-assisted XR system to address the above challenges. First, a mobility-aware resource provisioning scheme is proposed to enhance resource utilization while satisfying user QoE on a large timescale. Specifically, we present a mobility model tailored for XR users to capture both user spatial movements and interaction features. Then, we estimate user-specific model parameters and adopt a sample average approximation method to model the relationship between user QoE and the consumption of communication and computing resources. A coordinate descent algorithm is designed to make resource reservation decisions, where a deep neural network provides a valuable initial point to accelerate convergence. Simulation results demonstrate that the proposed resource provisioning scheme is more efficient in reducing network resource consumption while satisfying user QoE, compared with benchmark schemes. Second, we develop an adaptive volumetric video caching and rendering scheme to enhance real-time user QoE by considering dynamic user viewing behaviors. Particularly, volumetric videos of different quality levels need to be cached, rendered, and delivered to XR devices for different viewing distances within a time latency. Given limited resources for the service, we formulate a user QoE maximization problem to jointly optimize volumetric video caching and rendering decisions based on users’ real-time locations and viewing distances. To solve this problem, we first design an online regularization-based optimization algorithm to obtain caching decisions. We then present a low-complexity binary search algorithm to determine optimal rendering quality. Simulation results demonstrate that the proposed scheme achieves higher real-time user QoE in comparison with benchmark schemes. Third, we design a scheme for joint selection of rendering quality and encoding type by considering the interdependency among edge processing tasks to enhance long term user QoE. To cope with network dynamics, the rendering quality of frames can be dynamically adjusted, which in turn triggers an intra-frame encoding and leads to a sudden transmission burst. To capture such task interdependency, we formulate a long-term QoE maximization problem under edge computing and communication resource constraints, which jointly selects the rendering quality and either intra- or inter-frame encoding for each frame. To solve this problem, we theoretically analyze the impact of per-frame decisions on long-term QoE and present an online algorithm for decision-making. Simulation results demonstrate that the proposed joint rendering quality and encoding type selection scheme can further enhance resource utilization and long-term user QoE compared with benchmark schemes.

In summary, we have proposed a mobility-aware resource provision scheme, an adaptive volumetric video caching and rendering scheme, and a task dependency-aware rendering quality and encoding type selection scheme for an edge-assisted XR system. This research should provide useful insights for network operators to deliver immersive XR services at low operational costs.