Incorporating stochasticity in demands for optimizing resource allocation in versatile edge systems devoid of layer constraints

Abstract

While stochastic demand models have been introduced to better capture resource fluctuations, each instance of problem under different architectures needs to be analyzed and solved separately with duplicated work, this limitation motivates research to develop a scalable and adaptive resource scheduling solution capable of handling edge systems with any number of layers. Fortunately, we found that this kind of stochastic scheduling problem combined with sharing has an invariant optimal substructure that is independent of the number of layers, and propose the Placement of Resource in Any layer Edge architecture (PRAE) algorithm. We find an efficient way to solve it by splitting it into multiple subproblem groups and characterizing the relationship between these subproblems as an isomorphic network. We then identify optimal conditions for subproblem groups at different layers, and quickly achieve a solution using dynamic programming. Extensive evaluations show that PRAE improves resource utilization by over 28% compared to average demand models and achieves over 98% of optimal performance across diverse scenarios, with significantly lower computational complexity.