Latency-sensitive edge-native applications may be the key to commercial success of edge infrastructure. However, success in the form of widespread deployment of such applications poses its own challenges. These applications are edge-dependent by definition, and therefore cannot simply fail over to the cloud if the edge is overloaded. In this paper, we propose an adaptation-based strategy to allow scaling up the number of concurrent edge-native applications on a resource-limited cloudlet and wireless network. We demonstrate up to 40% reduction in offered load with minimal impact on latency on a variety of cognitive assistance tasks over non-adaptive approaches. Our approach is able to gracefully degrade and maintain quality of service for a subset of applications in the face of severely loaded conditions.

Wang, J, Feng, Z., George, S., Iyengar, R., Pillai, P., Satyanarayanan, M.George, S., Wang, J., Bala, M., Eiszler, T., Pillai, P., Satyanarayanan, M.
Proceedings of the Fourth IEEE/ACM Symposium on Edge Computing (SEC 2019), Washington, DC, November 2019