Edge computing is becoming a complementary architecture to cloud computing. A key advantage of edge computing is that edge data centers are typically located much closer to users than traditional cloud data centers. This is especially relevant for applications with tight latency requirements, like virtual reality, autonomous vehicle or the Internet of Things. However, since edge clusters are often resource-constrained, there is a trade-off between network latency and server processing latency while deploying applications on both edge and cloud.

The authors propose an approach aiming at devising hybrid cloud/edge deployment plans guaranteeing an end-to-end latency (i.e., mean response time) lower than that of the pure cloud deployment counterparts, at minimal cost. For a given scenario with number and location of users and data centers, the method optimizes the number of servers per data center as well as the probability, for each possible pair user / data center, that a request from the user is scheduled to the data center. The resulting probability matrix enables the scheduler to realize the trade-off between edge processing and cloud processing.

Because the mean end-to-end response time is a highly nonlinear and nonconvex function of the number of servers and scheduling probabilities, the authors used the Interior-Point/Direct algorithm of Artelys Knitro with the multistart feature to solve a continuous relaxation of the problem before rounding. Results show that the proposed approach indeed produces deployment plans at least as efficient as cloud deployments for an equal or lower budget. In addition, the proposed method outperforms purely edge deployments (i.e., when users send their requests to the closest data center) by 11% to 24%.