Enabling multi-hop remote method invocation in device-to-device networks

Resource sharing or computation offloading on mobile networks can bring a lot of benefits, the approaches in this domain if possible, can be widely applied to IoT networks or ubiquitous cities. In this section, we will go through the several limitations of current technologies, as well as our motivation to build a middleware that overcomes these obstacles by extending Remote Method Invocation method and multi-hop capability to enable resource sharing over mobile networks.

Low-end devices always have trouble running intensive resource-consuming applications such as image or video processing, which remarkably slow down its speed and drain energy. One well-known solution is having the device participate in a collaboration in which it can offload or migrate intensive code portions [1‐3] onto another device or cloud server with copious resource capacity, have them execute the code and wait for responses [4, 5]. Although the idea is straightforward, code offloading has not been widely applied in the mobile industry due to several issues: it requires radical changes to the system core (e.g. OS rooting) and originates high latency, making it inapplicable for the category of applications that send a huge number of requests in a short period (e.g. real-time).

Remote Method Invocation (RMI) is a distributed architecture in which methods of remote Java objects can be invoked from other Java virtual machines possibly located on different hosts [6]. One of its strengths is that it provides a degree of location transparency by having servers add services to a registry and requiring clients to use the registry for binding. Having RMI enabled on mobile platforms can bring several benefits, which the best of those is device resource transparency when multiple resources can be allocated for one function call. However, the original RMI technology does not support any routing other than what the underlying network layers support, if there is no inter-network (network-layer routing) between two devices, an object on the first device cannot invoke a method for an object on the second device. Moreover, RMI is an obsolete technology that heavily relies on the javax package from the Java SDK library; therefore, it is not supported on mobile platforms like Android.

Technology based on object brokers, like CORBA [7], can support remote method invocations. However, they rely on middleware processes (or threads) to instantiate or re-hydrate objects and then bind method calls to the target objects. Although there have been attempts to support CORBA on mobile platforms [8], they require the underlying layers to handle inter-networking and therefore do not directly support inter-group communications. In addition, the authors believe that its middleware is too heavy for most mobile devices and that its language-neutral approach to distribution is unnecessarily complex for most mobile apps.

RabbitMQ [9] or ActiveMQ [10] or the other current message queues are the fully implemented middleware that is widely used in many different research and commercial products; they can be deployed in distributed and federated configurations to support high-scale and high-availability software architectures. However, the main obstacle of these popular middleware systems is that the central server application must be located on a stationary server not a mobile device because it requires considerable system resources and platform dependencies [11, 12], and thus the entire system is unable to deploy on a self-operating mobile network. Likewise, the same problem also occurs for other publish-subscribe middleware systems in which a developer needs to become familiar with the libraries that always overwhelm the mobile platforms.

Most of the new generation phones feature closed-range, non-Internet communications such as Bluetooth, NFC and WiFi-Direct [13]. However, while WiFi-Direct can allow handshake between two devices that are nearly 200 m apart, Bluetooth and NFC can only work at small distances. While Bluetooth and NFC can only pair between two devices, WiFi-Direct enables connections between an unlimited number of devices as long as they are within the supported distance range. A WiFi-Direct network is a client-server model in which one device is elected to become a group owner, and the others connect to the owner as the clients. According to this model, once a device joined a group, it is by default unable to be contacted by any other groups. There are some solutions to address this limitation [5], but, these solutions lack a standard software model or out-of-the-box library to quickly develop or integrate into a software system.

To simplify the networking development of mobile applications and extend the limited range of non-Internet communications such as WiFi-Direct, we introduce our new middleware system that enables remote method execution routing on device-to-device networks (D2D). Our library adopts a group-to-group network architecture that can extend the range of communication, so that a device from one group can talk to a device from another group through a virtual bridge. Firstly, the developer implements functions and declares annotations on those he/she wants to publish to define the service. Then, during the code compilation, the compiler will automatically generate the software components that will be used to construct D2D networks. By integrating these constituent components into the app, a user can initiate a network in any topology since these components can flexibly switch between the devices. In “Evaluation” section we will demonstrate the use of the software library to quickly build two applications, chat and remote browser, that enable communications over multi-group device-to-device network. In this article, we introduce our new middleware architecture that makes the following contributions:

Although the research aims to extend the communication capacities of mobile devices in all cases, this article only focuses on one network interface that is WiFi-Direct because of its long range distance and availability (whenever a Wi-Fi network is active). We believe the same system architecture can also be applied to the other network interfaces on the same device such as NFC or Bluetooth.

Our system performs RMI by serializing a function call into a binary stream and dispatching over a wireless network. In the same domain, Android RMI [20] leverages the original Binder to allow users to invoke system services as well as application services between devices using a remote parcel format. Lin et al. introduces a cross-platform IPC mechanism called XBinder [21] to enable remote processes among multi-user communication for mobile applications to cooperate with local or remote services without forming a complicate network. However, despite the remarkable improvements with respect to performance [22], these design choices only target mobile devices connected in the same network, and thus they will not work when devices are moving to the other networks. To address this issue, we adopted a group-to-group network architecture to help flexibly switch the roles of devices during run-time and allow reconfiguring the network in multiple topologies.

Nakao et al. [23] provides an RPC-based invocation mechanism between Android devices using Intent, a message format used by the Android platform to realize transparent remote service communications to other devices without any modifications to the existing Android applications. Similarly, Nagahara et al. [24] proposed a distributed intent framework in which Android applications collaborate with embedded devices by sending serialized Intent messages through the network. Another approach for mobile remote processes is making services public, so that other devices may invoke services using remote call mechanisms [25, 26], but this approach incurs too much overhead for the host device as well as posing risks of unavailability of services when the host moves out of the communication range.

Our middleware contributes to the domain of WiFi-Direct multi-group communication in which one group connects to another using a legacy client and a special component operating on the original Wi-Fi interface to serve as a bridge between the two group owners [14, 27, 28]. Casetti et al. [17] leverages Wi-Fi Direct to support multiple groups for a content-centric routing network in which data is transparently available to users using content routing tables that collect and transport data over the content nodes. Before the execution, routing tables are advertised and populated via a registration/advertisement protocol. Our system extends the idea of the content-centric network to bring the function-centric mobile network, in which any device can request for a function call regardless of knowing the actual location of the function (i.e., the requested function can be hosted on a certain mobile device or a stationary server).

In the category of the wireless P2P communication before the Wi-Fi Direct technology, several efforts utilized wireless communications in an add-hoc fashion to establish P2P networks, such as media sharing systems on urban transportation using Bluetooth [29], resource sharing using cellular networks [30], and radio resource sharing over ultra-wideband [31]. Built on top of Wi-Fi P2P, Rio [32] leverages I/O devices to capture and share contents and resources between the existing applications running on different devices without any modification. Some of its applications are multi-system photography and gaming, music and video sharing, and SIM card sharing for multiple devices. GameOn [33] was also built on the same network infrastructure to establish non-Internet connection between gamers within closed range networks like on public transportation. CAMEO [34], and GigaSight [35] are also the similar content sharing systems in closed range network architectures.

Finally, although we used real devices and environments for our experiments, there is a large gap between testing devices and the real world that contains hundreds of devices and different situations. Therefore, we share the same vision with the network simulation research, especially for multi-group WiFi-Direct networks to overcome following two issues: (1) the high cost of the experiment deployment with a vast number of devices and (2) the complexity of the network discovery and handshake phases. WiDiSi [6] is a dedicated visual simulation extending the PeerSim library [36] to support WiFi-Direct, it can simulate and visualize a vast D2D network including the discovery and network establishment of devices moving randomly within closed distances. However, the disadvantages of WiDiSi as well as PeerSim are being single-threaded, having less autonomy and unsupported not supporting multiple groups. The new result of that work, called MAGNET [8], is a novel self-organizing middleware infrastructure that aims to provide reliable and stable P2P connectivity among large numbers of smart devices.

We built our middleware on top of the ZeroMQ (ZMQ) library [37], a flexible library for message queues which is available on multiple platforms including Android. As a result, our first version works on both Android and PC systems. Our middleware system can be simply employed via two steps: a developer creates the service with full implementation and marks our service annotations (Code Snippet 4). After compiling the project, the middleware’s processor will automatically generate extension classes for the service; then, the developer uses these classes along with the other basic components such as Broker and Bridge to construct their mobile networks (Code Snippet 5).

Since the middleware system addresses the issue of extending the limited communication range of mobile networks, especially when there is no need of Internet, it can be used across many different applications. In this section, to show the ease of a software development, we further discuss two mobile applications that utilize the proposed middleware system.

For the evaluation, we built a testbed with Wi-Fi Direct featured by Android devices to evaluate the performance of the developed middleware system. A personal computer is also included to examine the bridge between mobile devices and stationary computers (Table 1).

In this article, we introduced a new middleware architecture to enable routing remote method invocation over multiple group device-to-device networks. Our work modularizes its architecture by the functional components using annotations, which makes it flexible to apply and adaptive with either D2D or device-to-server networks. Our case-study applications and empirical experiments tested on the actual testbed and real mobile devices unveil the ease of using the library in mobile software development, the low overhead conduction and high performance. In the future, we will consider these following directions (1) optimize the selection algorithm at Broker side to fairly distribute a request to a number of Workers for higher availability of mobile execution resources and better performance. (2) Secondly, to aim for large-scaled IoT networks, the middleware must comply with multiple criteria and test scenarios from a huge number of devices and different specifications, as well as rapid changes of device location and environment. This requirement can be resolved by a simulation where device specs and mobility are simulated in a predefined environment with several prerequisite conditions. Finally, (3) we will extend middleware library with respect to streamlining and easiness to target a larger scope of applications, by simplifying the APIs to reduce the cost of integration and development.