An effective offloading middleware for pervasive services on mobile devices

Abstract

The practical success of pervasive services running in mobile wireless networks and devices relies on their ability to provide effective and efficient offloading support, so as to satisfy the increasing demand for mobile devices to run heavier applications (e.g. those running on desktop PCs). Offloading is an effective mechanism for leveraging the severity of resource constrained mobile devices by migrating some computing load to nearby resource-rich surrogates (e.g. desktop PCs, servers) on home networks or their extension. This paper proposes a light-weight and efficient offloading middleware, which provides runtime offloading services for resource constrained mobile devices. The middleware considers multiple types of resources (i.e. memory, CPU and bandwidth) and carries out application partitioning and partition offloading in an adaptive and efficient manner. The corresponding algorithms are presented. The evaluation outcomes indicate the effectiveness and efficiency of this service offloading solution.

Introduction

Recent years have seen a significant increase in research and development of mobile and wireless networks, including UMTS (Universal Mobile Telecommunications System), IEEE 802.11/16/20, mobile ad hoc networks, and sensor networks. The future success of these network systems lies in their ability to provide users with cost-effective services that have the potential to run anywhere, anytime and on any device without (or with little) user attention. Services with these features are termed pervasive services, which is an active branch of pervasive (or ubiquitous) computing [1].

With advances in mobile terminals and wireless communications, the demand for mobile devices to run heavier applications (such as video playing or editing) is increasing. Apparently, even the most powerful PDAs today (due to the size and weight constraints) are unable to compete against their desktop siblings with regard to any type of resource, especially battery life and network capacity. Meanwhile, in some working, entertaining and living environments, computing resources are often rich. For instance, in offices or cafes, some desktop PCs may be idle while mobile devices are busy. As such, it makes sense for these resource constrained mobile devices to make use of the resources available in their vicinity to leverage their resource insufficiency. For example, it is natural for John to expect to watch the video clips of his daughter’s Christmas concert (even if with degraded quality), which was just uploaded to their home server, during a tedious wait in a Heathrow boarding lounge. Therefore, there is a clear need for mobile devices to run heavier applications. We refer to this kind of computation or communication migration as offloading for pervasive services. It is also called surrogate computing or cyber foraging [10], [1].

Offloading is not a new concept. It has been around in the form of proxies and surrogates for many years. For example, it has been used for load balancing in distributed systems [1]. However utilizing offloading mechanisms in the domain of resource constrained mobile devices has gained popularity lately.

Since pervasive devices are heterogeneous in terms of CPU power, memory, communication capabilities, battery life and software features, middleware is the most common solution for facilitating interoperability in pervasive environments [2]. The aim of our work is to provide a light-weight and effective middleware for pervasive service offloading in mobile devices.

Some research work on offloading mechanisms for resource constrained mobile devices has been proposed [3], [4], [5], [6], [7], [8], [9], [10], [11]. However, they are limited in one way or another. Since pervasive services tend to be more complex and dynamic [3], ideally, an offloading system should consider multiple types of resource constraint. In this paper, we present our work on a runtime offloading middleware specifically for pervasive services. Instead of considering only one type of individual resource, such as memory [3] or CPU [4], our offloading approach counts on a combination of these plus communication cost (i.e. bandwidth resources). Through grasp of service runtime feature in a cost-effective manner, the efficiency of our offloading approach is significantly increased. The adaptability and efficiency of our offloading middleware also lies in its light-weighted service partitioning algorithm and feasible offloading mechanisms.

As far as middleware for pervasive computing is concerned, the Spectra project [5], [9] proposes a remote execution system for mobile devices used in pervasive computing. Chen et al. proposed an offloading framework [4] that focuses on a Java-based environment and dynamically decides whether to execute locally or remotely, based on the cost of Java-code compilation, computational complexity and communication channel conditions. The Coign [6] project proposed a system to use a min-cut algorithm to statically partition binary applications built from Microsoft’s Component Object Model (COM) components. Li et al. [7] constructed a static cost graph and applied a partition scheme to statically divide an application’s tasks into client and server subtasks during the design time. Goyal et al. [11] proposed a cyber foraging solution using virtual machine technologies. However, these works did not provide specific offloading mechanism.

Gu et al. [8] proposed an adaptive infrastructure for Java application offloading execution, which adapted a min-cut [12] heuristic algorithm to dynamically partition an application. In addition, the Gu et al. work considers only memory and supports only one surrogate, however, most mobile applications are more sensitive to bandwidth and CPU cycles, as is the case in the John’s video playback scenario above. As such, an ideal partitioning solution should consider memory, CPU and bandwidth simultaneously. Focusing on only one of them will render it unfeasible in practice. Our offloading middleware considers both the interaction properties and the resource consumption when conducting partitioning and offloading. And we aim to relieve not only memory constraints but also CPU usage and bandwidth constraints.

The remainder of this paper is organized as follows. Section 2 describes a scenario for offloading services and presents some issues involved in designing offloading systems. The architecture of our proposed offloading middleware is discussed in Section 3. In Section 4, we present a class instrumenting approach for offloading purposes. A Service partitioning algorithm is presented in Section 5. In Section 6, we discuss some practical issues for implementing the offloading middleware. Implementation and evaluation comprise Section 7 and the paper concludes in Section 8.