Smoothed Online Resource Allocation in Multi-Tier Distributed Cloud Networks

Abstract

The problem of dynamic resource allocation for service provisioning in multi-tier distributed clouds is particularly challenging due to the coexistence of several factors: the need for joint allocation of cloud and network resources, the need for online decision-making under time-varying service demands and resource prices, and the reconfiguration cost associated with changing resource allocation decisions. We study this problem from an online optimization perspective to address all these challenges. We design an online algorithm that decouples the original offline problem over time by constructing a series of regularized subproblems, solvable at each corresponding time slot using the output of the previous time slot. We prove that, without prediction beyond the current time slot, our algorithm achieves a parameterized competitive ratio for arbitrarily dynamic workloads and resource prices. If prediction is available, we demonstrate that existing prediction-based control algorithms lack worst case performance guarantees for our problem, and we design two novel predictive control algorithms that inherit the theoretical guarantees of our online algorithm, while exhibiting improved practical performance. We conduct evaluations in a variety of settings based on real-world dynamic inputs and show that, without prediction, our online algorithm achieves up to nine times total cost reduction compared with the sequence of greedy one-shot optimizations and at most three times the offline optimum; with moderate predictions, our control algorithms can achieve two times total cost reduction compared with existing prediction-based algorithms.