Learning-Based Localized Offloading with Resource-Constrained Data Centers

Offloading has emerged as a new paradigm to save energy for mobile devices in the context of cloud computing systems. Unlike the traditional cloud computing, it offers the flexibility of switching between local and remote execution, and employs accurate profiling of tasks. Given a resource-constrained data center, an interesting optimization question is which tasks should be offloaded/run locally so that global energy savings is maximized. The main technical difficulties are related to the uncertainty and variability of congestion, as well as the need for a real-time, low overhead and localized decision procedure that are near optimal. We introduce a combination of statistical and learning-based techniques that use the results of offline centralized algorithms to create localized online solutions that perform well under realistic workloads. The procedures and algorithms are compared with upper bounds to demonstrate their effectiveness.