No abstract

This paper proposes and evaluates MALMOS, a novel framework for mobile offloading scheduling based on online machine learning techniques. In contrast to previous works, which rely on application-dependent parameters or predefined static scheduling policies, MALMOS provides an online training mechanism for the machine learning-based runtime scheduler such that it supports a flexible policy that dynamically adapts scheduling decisions based on the observation of previous offloading decisions and their correctness. To demonstrate its practical applicability, we integrated MALMOS with an existing Java-based, offloading-capable code refactoring framework, DPartner. Using this integration, we performed quantitative experiments to evaluate the performance and cost for three machine learning algorithms: instance-based learning, perceptron, and naïve Bayes, with respect to classifier training time, classification time, and scheduling accuracy. Particularly, we examined the adaptability of MALMOS to various network conditions and computing capabilities of remote resources by comparing the scheduling accuracy with two static scheduling cases: threshold-based and linear equation-based scheduling policies. Our evaluation uses an Android-based prototype for experiments, and considers benchmarks with different computation/communication characteristics, and different computing capabilities of remote resources. The evaluation shows that MALMOS achieves 10.9%∼40.5% higher scheduling accuracy than two static scheduling policies.

No abstract

No abstract