Recent advancements in 5G technology have enabled significant breakthroughs in applications requiring high data throughput and low latency, such as Mobile Augmented Reality (MAR). Nevertheless, mobile devices tend to be computationally and energy-constrained, limiting their capability to provide synchronous interactions between the real environment and visual augmentations. Thus, in this work, we propose a collaborative inference method to run compute-intensive AR tasks, such as object recognition, on heterogeneous devices at the network edge. This approach would benefit both end-users and service providers by reducing end-to-end latency and operating costs, respectively.
This paper analyzes the performance of disaggregating monolithic applications across heterogeneous hardware. A monolithic program with convolution functions was tested on AMD, Intel servers, and a BlueField-2 DPU (ARM). After disaggregation into independent processes, it was run on the same server and over a local network. Results showed that disaggregation maintained similar performance to the monolithic version, with minimal impact despite network latency. We also observed improved cache utilization, particularly a reduction in L1 cache misses and L2 accesses. Disaggregation shows promise in optimizing memory usage. Future work will focus on refining memory analysis and using GPUs and FPGAs to improve computational efficiency.