MOBA: Real-Time Model Selection and Bandwidth Allocation for Wireless Image Transmission

We propose MOBA, a real-time scheduler for MOdel selection and Bandwidth Allocation, to provide low-latency and high-quality image transmission in wireless networks. Specifically, MOBA first decomposes the joint optimization into large numbers of simple subproblems that are suitable for parallel computation. Then, we can solve them in parallel through sampling a subset of subproblems from identified promising sets, achieving real-time near-optimal solutions. Finally, we implement MOBA on an off-the-shelf Graphics Processing Unit (GPU) platform. Experimental results show that MOBA outperforms other traditional schemes, while meeting the stringent timing requirement (125 μs) in 5G.