Smoothed Dynamic Allocation of Multi-Tier Distributed Capacities
In the emerging edge computing paradigm, the edge clouds, small-scale and highly distributed, are introduced into the service path between customers and the conventional and gigantic clouds at the Internet core, forming a tiered structure of resource pools. An important issue here is the joint resource allocation at the clouds of multiple tiers and at the network connecting them, complicated by the challenges of the dynamic and often unpredictable workload, the reconfiguration of resource allocations across consecutive time slots, as well as the cloud and the network heterogeneity which may also be time-varying, the geographic distribution which constrains the SLA, etc. We study this multi-tier resource allocation and reconfiguration problem from an online optimization perspective, capturing all the above challenges. To overcome the fundamental difficulty of the reconfiguration that couples the resource allocation decisions made at consecutive time slots, we carefully construct a different yet related problem decoupled and solvable at every time slot, "regularizing" the decisions to a bounded, exponential decay from a geometric point of view, and we formally prove that, without any lookahead or prediction beyond the current time slot, using the solutions to this constructed problem as the solutions to our original problem can provide a parameterized competitive ratio for any arbitrary workload. We extend such an optimality guarantee to arbitrary N (N >= 2) tiers of clouds. Evaluations in a variety of settings based on a large number of clouds with real-world workloads of regular and flash crowd fluctuations confirm that our online algorithm performs well in practice, achieving up to about 80% less total cost across time than the one-shot optimization and up to about 3 times the offline optimum in a common 2-tier edge computing structure.