Machine Learning-Based Runtime Scheduler for Mobile Offloading Framework

Remote offloading techniques have been proposed to overcome the limited resources of mobile platforms by leveraging external powerful resources such as personal work-stations or cloud servers. Prior studies have primarily focused on core mechanisms for offloading. Yet, adaptive scheduling in such systems is important because offloading effectiveness can be influenced by varying network conditions, workload requirements, and load at the target device. In this paper, we present a study on the feasibility of applying machine learning techniques to address the adaptive scheduling problem in mobile offloading framework. The study considers 19 different machine learning algorithms and four workloads, with a dataset obtained through the deployment of an Android-based remote offloading framework prototype on actual mobile and cloud resources. From this set, a subset of machine learning algorithms, which have relatively high scheduling accuracy, is selected to implement an offline offloading scheduler. Finally, by taking computational cost and the scheduling performance into account, we use Instance-Based Learning to evaluate an online adaptive scheduler for mobile offloading. In our evaluation, we observe that an Instance Learning-based online offloading scheduler selects the best scheduling decision in 87.5% instances, in an experiment setup in which an image processing workload is offloaded while subject to varying network bandwidth conditions and the amount of data transfer.

I. Introduction

Rapid enhancements in computing capabilities of mobile platforms have been driving the increased adopting and use of mobile computing platforms by increasing numbers of users. Today's mobile platforms are able to deliver capabilities that are close to those of non-mobile platforms such as desktops or workstations. For instance, a mobile phone equipped with a Graphic Processing Unit (GPU) core is able to achieve approximately 10GFLOPS/Watt of computer-power, which is identical as a 4-core desktop with GPU [1]. Despite of these significant advancements, mobile platforms remain significantly limited by resources such as memory size, storage capacity, and especially battery lifespan. To alleviate the problem of the resource limitations in mobile platforms, computation offloading techniques have been proposed as a way to extend the capabilities of mobile platforms to more powerful resources. These may include personal computers, servers, or even public cloud resource over the network [2], [3], [4].