Smoothed Online Resource Allocation in Multi-Tier Distributed Cloud Networks

In the emerging edge computing paradigm, smallscale highly distributed edge clouds are on the service path between end users and conventional large-scale clouds at the Internet core. A crucial problem that needs to be addressed in order to drive cost and performance in this multi-tier distributed infrastructure is the dynamic and joint allocation of cloud and network resources, which is particularly challenging due to the coexistence of several factors: the reconfiguration cost associated to changing resource allocation decisions over time, the constantly varying and often unpredictable nature of service demands, as well as the heterogeneity of distributed resources. We study the problem of resource allocation and reconfiguration in the multi-tier resource pool from an online optimization perspective that addresses all the challenges above. Our approach decouples the original problem over time by constructing a series of subproblems that are solvable at each corresponding time slot using the output of the previous time slot. Via solid formal analysis, we prove that, without any lookahead beyond the current time slot, our online algorithm provides a solution with a parameterized competitive ratio for any arbitrarily dynamic workload and operating price. We conduct extensive evaluations in a variety of settings based on a number of clouds and real-world workloads with regular and flash crowd fluctuations, and demonstrate that our online algorithm performs well in practice, achieving up to 9x total cost reduction than the sequence of one-shot optimizations and at most 3x the offline optimum.