Collaborative computing system for heavy processing and low latency in wireless sensor networks

HeaLow: A Collaborative Computing System for Heavy Processing and Low Latency in Wireless Sensor Networks

Introduction

Supported by advancements in Microelectromechanical Systems (MEMS), sensor technologies have progressed significantly, with various sensor types emerging [1]. With such progress, Wireless Sensor Networks (WSNs) have attracted considerable attention from researchers in diverse fields, including the Internet of Things (IoT) [2, 3]. WSNs comprise interconnected sensor nodes, with sensor nodes responsible for collecting sensed data and transmitting it to users or servers [4, 5, 6]. Generally, wireless sensor nodes have weaker resources compared to wired nodes, as wired power sources are typically avoided due to mobility concerns [7]. Therefore, existing WSNs are mainly used for delay-tolerant services requiring low computation. For this reason, many researchers have focused on improving the power consumption efficiency of wireless networks [8, 9]. Furthermore, other researchers have conducted studies on using cloud servers to overcome the limited resources of nodes, such as low computing power, storage scarcity, etc. [10, 11, 12, 13, 14]. This is because tasks with delay tolerance or wide-area aggregation features are well-suited to the powerful resources of cloud servers [15].

Considering the advancement of hardware and software technologies in recent years, today’s mobile devices have more resources, sensors, and actuators. In addition to mobile devices, there have been significant advancements in small and low-cost sensor platforms such as Arduino and WRTnode [16, 17]. Using these devices, users can collect various sensor data and perform the required heavy computational processing. Based on such progress, the concept of edge computing was proposed. Edge computing refers to networks where edge nodes actively participate in computational operations [18]. Existing cloud computing models are mainly designed for traditional web applications, so they are not suitable for future internet applications running on various mobile and sensor nodes [19]. Moreover, network latency caused by long distances between cloud servers and devices is unsuitable for delay-sensitive applications or services that require frequent transfer of large-sized data, such as video and audio data [16]. Conversely, edge computing allows sensor nodes to properly use their resources for processing. Therefore, a wider range of applications can be executed on sensor nodes with the concept of edge computing.  

As mentioned above, today’s nodes used for WSNs have significantly more resources than in the past. However, performing heavy computations and low-latency processing is still difficult for them. Therefore, in addition to edge computing, collaborative computing was proposed to use sensor node resources more effectively. In collaborative computing, sensor nodes share their resources, such as computing power and energy, and perform processing collaboratively. Based on collaborative computing, sensor nodes can offload the computation of sensed data processing to improve the computational performance and energy consumption of sensor nodes uniformly. Therefore, nodes using collaborative computing can execute applications that require large computations. In particular, most applications using video data require processing a large amount of computation. Therefore, the performance of such applications can be improved by applying the concept of collaborative computing for applications in WSNs. For example, there is research that proposes a system for using collaborative video processing to improve the quality of the processing result [21].

In addition to video data, audio data is also one of the most widely used data types in WSNs. Microphones are inexpensive and widely used, so most mobile devices already have a built-in microphone. Furthermore, even if some sensor nodes do not have a built-in microphone, most of them are capable of using a microphone. This means that applications using audio data can easily use nearby devices as sensor nodes. Due to these advantages, many researchers have been interested in using audio data in WSNs, and various applications have been proposed using audio data [22, 23, 24, 25, 26]. For processing or analyzing audio data in such applications, Fast Fourier Transform (FFT) [27] is almost always a key component. Even cross-correlation [28] or wavelet transform [29] are basically a combination of several FFT operations or a partial result of FFT. Since FFT requires a lot of computational resources, audio data processing operations require a large amount of computation. Moreover, many applications require that the operations be performed on time, which makes it difficult for a node in a WSN to execute relatively heavy applications.

To overcome the limitations, we propose a collaborative computing system, HeaLow, for heavy processing and low latency in WSNs. To complete heavy computations within the deadline, HeaLow performs processing and offloading properly by considering various factors. Since audio data processing (audio signal processing) is one of the most widely used operations for applications in WSNs, we consider audio data processing as the target task in this paper. However, HeaLow can be used for any application that requires large computations and processing to be completed within a limited deadline. The design goal of HeaLow is to help incapable wireless sensor nodes complete heavy computations within the deadline. Furthermore, if sufficient resources are available in the multi-device system, using free resources to improve application performance is HeaLow’s secondary goal. Since the timeliness of heavy computations is a challenging task, the issue of power efficiency is not considered in this paper.

In this paper, by leveraging the powerful resources of today’s devices, we focus on using several more powerful nodes for low-power devices and propose a system that enables incapable nodes to complete heavy computations within the deadline, which is new and different from traditional offloading research, including those on WSNs. The contributions of this paper are summarized as follows:

• Using HeaLow, nodes in a WSN can perform heavy computational processes while meeting the completion time requirement.
• Unlike related studies that perform evaluations only using simulation, we implemented HeaLow on real devices. Therefore, we evaluated HeaLow not only through simulation but also through several experiments using devices.
• In this paper, we consider audio signal processing, which can be used for various services and applications in WSNs, as the target task. Therefore, such services and applications can easily use HeaLow.

The rest of the paper is organized as follows. First, we introduce the related work and describe the new advantages and benefits of HeaLow compared to related works in Section 2. Section 3 describes the design and implementation of HeaLow. Section 4 explains the experiments and simulations and confirms the performance of HeaLow. Finally, Section 5 concludes this paper.

Related Work

In this section, we introduce studies on cloud-assisted WSNs, task offloading in edge computing, and collaborative computing. In addition, we introduce some recent studies on audio sensor-based applications. After that, we show the limitations of the related work and the differences between our research and theirs.

Generally, nodes in WSNs have relatively limited resources, so many studies have been conducted to overcome the limitation by using the powerful resources of cloud servers. First, Wan et al. focused on energy-optimal application execution in their cloud-assisted platform [30]. To reduce the energy consumption of sensor nodes, they developed strategies for clock frequency configuration and optimal data transmission scheduling. According to their numerical results, a large amount of energy can be saved by offloading the task to a cloud clone. Similar to Wan [30], Sinha et al. used server resources to reduce device limitations, but they used multiple servers together based on their task partitioning algorithm [31]. Their evaluation results show that the proposed partitioning algorithm performs better than existing algorithms. Unlike Wan [30] and Sinha [31], Dattatraya et al. focused on the insufficient storage of sensor nodes [32]. In their system, sensor nodes store sensed data in servers, and the system provides access to authorized users to retrieve sensor information stored in the servers. Among the above work, many studies have shown that using cloud servers is an effective way to alleviate the resource limitations of sensor nodes. However, according to Chandra [16], using cloud servers is unsuitable for delay-sensitive applications or services that require frequent transfer of large-sized data. Moreover, existing cloud computing models are mainly designed for traditional web applications, so these models are not suitable for future internet applications running on various mobile and sensor nodes, including the work done by Sheng et al. [19].  

In recent years, the devices of sensor nodes in wireless networks have made significant progress, so various processes can be performed in sensor nodes without a cloud server. Therefore, many researchers, such as Varghese et al., focused on edge computing and conducted studies on task offloading, which is an important process in edge computing [33]. First, Kyung et al. proposed a method for optimal partitioning among nodes to minimize the search time of the pattern of interest in a large data [34]. They performed a simulation evaluation, and the simulation result shows that the expected processing time is reduced by dividing the tasks among the nodes. Chen et al. designed a distributed computational offloading algorithm and formulated the offloading decision-making problem [35]. They conducted numerical studies, and the results of the studies show that their algorithm improves computational offloading performance and scales well depending on the number of nodes. Similar to Chen [35], Mao et al. proposed a dynamic computational offloading algorithm that jointly determines the offloading decision, CPU cycle frequencies, and transmission power [36]. The main advantage of this algorithm is that decisions can be made without requiring the distribution information of others. The proposed algorithm was evaluated through simulation.  

In addition to edge computing, some studies used the concept of collaborative computing to improve processing performance and uniform energy consumption of nodes. Sheng et al. proposed a new approach to minimize energy consumption in processing [19]. They evaluated the proposed solution through simulation, and the simulation results show that the total energy consumption.

SOURCE: www.mdpi.com