An Integrated Process-Network Load Balancing in Edge-Assisted Autonomous Vehicles Using Multimodal Applications With Shared Workloads

Recently, as computing-intensive services such as object recognition for autonomous vehicles have increased, power consumption and computational loads of vehicles have been prominent. To tackle this issue, there have been growing interests in edge computing technology, which offloads workloads of services to nearby vehicle edge computing (VEC) servers. However, the existing offloading technologies in the VEC server made independent offloading decisions for each service, and they did not consider shared workloads of multimodal vision applications. In this paper, we first aim to capture the intricate workload relationships among multimodal applications in modeling an integrated process-network load balancing for a VEC-assisted autonomous vehicle system. To this end, we formulate an energy minimization problem of a vehicle constrained by outage probability of service requests where the decision variables are (i) dynamic offloading policy between vehicle and VEC server and (ii) onboard CPU clock frequency of a vehicle every time slot. To solve this problem, we leverage Lyapunov optimization to transform the long-term average problem into a slot-by-slot problem. Then, by minimizing the slot-by-slot objective function every time slot, we develop a latency-sensitive energy minimization (LEMON) algorithm. Finally, we evaluate the performance of the proposed algorithm in realistic vehicular network environment, and show that the proposed LEMON algorithm which captures the shared workloads reduces 57% of average queue backlog and 37% of average power consumption compared to the existing algorithm which does not consider shared workload characteristics.