Performance prediction for supporting mobile applications offloading

Abstract

Mobile Cloud Computing (MCC) is the integration of mobile computing and cloud computing, and it can increase the performance of mobile apps and reducing their energy consumption through code and data offloading. Developers may build MCC systems on a public cloud. The public cloud may offer economies of scale, but there are some considerations to take into account. Cloud providers charge their customers by data traffic and use of virtual machines (VMs), and wrong offloading decisions may lead to financial losses. This dissertation proposes an approach for estimating applications’ performance, use of VM instances and data traffic generated by tasks offloading and its related costs on a public cloud. This work proposes two Stochastic Petri Net (SPN)-based formal modeling strategies to represent MCC applications and a cost model to predict data traffic volume and use of VM instances. The first SPN-based modeling strategy represents MCC applications running on user devices. The second one represents a remote infrastructure deployed in a public cloud for supporting offloading making by mobile users. By combining different instance types, simultaneous jobs per VM instance and thresholds for scaling the system, it is possible to offer different response times for each offloading scenario. In addition, using both strategies it is possible to represent the communication process between the app running on the user’s device and a remote infrastructure. Thus, making possible to estimate the performance of the MCC application. Our approach enables designers to plan and tune MCC architectures based on four performance metrics: Mean Time to Execute (MTTE), Mean Response Time (MRT), Cumulative Distribution Function (CDF) and Throughput. MTTE is related to the performance on the mobile device. On the other hand, MRT corresponds to the performance of the remote infrastructure deployed in a public cloud for supporting offloading. Our modeling strategy allows for the representation of the use and sharing of available bandwidth for offloading operations, as well as the effect of bandwidth variation on the metrics evaluated. It allows a more accurate evaluation by developers about the performance of their applications taking into account specific network requirements, users, and offloading scenarios. Four case studies were performed to evaluate our approach. Our approach has proven to be feasible and it highlights the most appropriate scenarios. Supporting developers at design time by providing statistical information about applications’ behavior and costs estimations. In addition, our approach may be adapted to support MCC applications in real time providing on-the-fly probabilistic performance predictions.