POSTER
Qinjun Jiang, Sarita Adve
Offloading compute-intensive tasks to edge or cloud servers enables high-performance, low-power interactive XR. However, existing distributed XR systems assume dedicated networks with stable low latency and high bandwidth, limiting deployment in shared and dynamic wireless environments. As XR moves beyond controlled local settings, network latency becomes a key performance bottleneck.
Modern networks provide heterogeneous paths with diverse latency–bandwidth tradeoffs, creating opportunities to improve end-to-end performance. Yet, the limited bandwidth of low-latency paths prevents uniform acceleration of all traffic, shifting the problem from how to offload to what to prioritize, where user-perceived quality depends on selectively accelerating critical traffic.
The key insight is that such prioritization requires integrating application- and runtime-level knowledge with network-level innovations. We demonstrate this through two case studies. For visual-inertial odometry (VIO) offloading, we explore how incorporating runtime knowledge such as user motion can better utilize low-latency paths compared to application-agnostic approaches. For rendering, we show that large message sizes and inter-frame dependencies hinder the use of low-latency paths. We therefore present a geometry-aided encoding pipeline that enables more flexible traffic steering and reprojection-aware optimizations to reduce bandwidth while preserving visual fidelity.
These results highlight the potential of cross-layer, application- and runtime-aware design for improving distributed XR over heterogeneous networks.