Traditional health care services can hardly meet the needs

Transcription

1 Mobile Big Data Fault-Tolerant Processing for ehealth Networks Kun Wang, Yun Shao, Lei Shu, Chunsheng Zhu, and Yan Zhang Abstract In daily life, people tend to use mobile networks for more accurate overall data. With intelligent mobile devices, almost all kinds of data can be collected automatically, which contributes directly to the blooming of ehealth. However, large amounts of data are also leading us into the era of big data, in which new data collection, transmission, and processing techniques are required. To ensure ubiquitous data collection, the scale of mobile ehealth networks has to be expanded. Also, networks will face more pressure to transmit large amounts of ehealth data. In addition, because the processing time increases with data volume, even powerful processors cannot always be regarded as efficient for big data. To solve these problems, in this article, an interests-based reduced variable neighborhood search (RVNS) queue architecture (IRQA) is proposed. In this three-layer architecture, a fault-tolerant mechanism based on interests matching is designed to ensure the completeness of ehealth data in the data gathering layer. Then the data integrating layer checks the accuracy of data, and also prepares for data processing. In the end, an RVNS queue is adopted for rapid data processing in the data analyzing layer. After processing with relevant rules, only valuable data will be reported to health care providers, which saves their effort to identify these data. Simulation shows that IRQA is steady and fast enough to process large amounts of data. Traditional health care services can hardly meet the needs of the growing population because hospital capacity and medical workers are limited in terms of the continuously increasing treatment requests. On this background, a new kind of ehealth service using intelligent device to monitor people s lives is developing rapidly with the benefits of big data technique. In this era of big data, large amount of data has been transmitted on the Internet, stored in servers and clouds, and even collected around people s life by mobile networks [1]. Characterized by their volume, velocity, variability, and veracity, mobile big data networks are used to describe those extremely large or complex data sets in the network for which traditional processing methods are inadequate. Specifically, data collected by mobile ehealth networks are becoming more ubiquitous with the development of hardware, and a group of collecting nodes are required to get more overall data [2]. Meanwhile, the number of collecting nodes in a network tends to increase, and each collecting node tends to Ken Wang is with Nanjing University of Posts and Telecommunications. Yun Shao is with the University of Southern California. Lei Shu is the corresponding author for this article, and is with Guangdong University of Petrochemical Technology, Dalian University of Technology, and Guangdong Provincial Key Laboratory of Petrochemical Equipment Fault Analysis. Chunsheng Zhu is with the University of British Columbia. Yan Zhang is with Simula Research Laboratory and the University of Oslo. have higher sampling frequency. Some typical types of sensors (e.g., brain sensors) will generate huge amounts of data, the size of which can even grow to the terabyte level. Also, the size of records of a clinic agency easily grows to the exabyte level. All the factors mentioned above increase the scale of a network, as well as the data volume transmitted in an ehealth network. On the other hand, due to the limited energy and functionality of mobile nodes, data has to be converged and processed in a central server. As shown in Fig. 1, the monitoring of people s status is always completed by the combination of body sensors and environmental sensors. The reason that sensor data (e.g., blood pressure, blood sugar, heart rate) from body sensors and environmental sensors should be combined in analysis is that some environmental factors (e.g., temperature, pressure) will affect an elderly person s physiological conditions a lot. In addition, we also show status sensors in Fig. 1. That is because moving status may cause inaccurate analysis. For example, the movement of some kinds of transportation (e.g., cars and elevators) will cause fall detection to be out of order in some cases. However, with the expansion of network scale caused by more powerful functionalities, it is highly possible that some interruptive nodes or network congestion block the routing so that it becomes difficult for data to reach the server. Accordingly, the server cannot have complete data for analysis, and it will be interested in the value of lost data. Based on this idea, the foundation of interests matching is formed: let a server generate interest requests according to lost data, and nodes match the received requests in their buffer; then the lost data could be recollected. In addition, on the server side, the increasing amount of data definitely slows down the processing speed. With the /16/$ IEEE IEEE Network January/February 216

2 IRQA Health care provider Congestion Status sensors Interruptive node Normal node Environmental sensors Brain Vision Hearing Blood pressure Falling Heart rate Toxins Body sensors Figure 1. Problem statement of mobile big data network. development of a network, it is highly possible that the processing ratio will be slower than the data receiving ratio, which means lots of the newest data arrives while old data is still being processed because of the increment of data processing time. Under this circumstance, if data is still organized in sequence as is tradition, a processor will analyze those earlier arriving data first, and the queue used to store the newest data is going to elongate infinitely so that the newest data will never be processed promptly. Therefore, how to process these collected data efficiently and return valuable information deserves further exploration. To this end, we mainly focus on reliable data transmission in a mobile ehealth network and rapid data processing in a central server under a big data scenario, and propose an interests-based Reduced Variable Neighborhood Search (RVNS) queue architecture (IRQA) to solve unreliable transmission and data processing problems caused by the expansion of a network. The contributions of our article are summarized as follows: A new fault-tolerant mechanism is proposed using interests matching to ensure the reliability of data transmission. An RVNS queue is designed based on the RVNS algorithm [3], which is always used for solving combinational optimization problems. RVNS queue improves processing speed dramatically under a big data scenario. Big data is a very hot topic, and much research is still in the exploration phase. Currently, research toward big data can be divided into two fields. Lots of research works focus on algorithms and strategies for big data mining. Zhang et al. designed a novel community-centric framework to predict community activities [4]. The framework consists of community detection and community activity modeling. It extracts community activity patterns from big data collected physically and virtually. Kuang et al. proposed a unified tensor model [5]. This model can represent unstructured, semistructured, and structured data with a tensor model. In detail, each kind of data is represented by a subtensor, which is finally merged with a unified tensor. Also, a small but valuable core tensor is extracted using an incremental high order singular value decomposition. Yu et al. designed an RTIC-C system to handle the huge data volume based on cloud computing [6]. RTIC-C includes a distribute data management service to ensure large-scale data storage, a parallel distributed framework to run various mining applications based on the Map-Reduce mechanism, and a restful web service interface for third-party mining services. Gu et al. proposed a cost minimization method [7]. In this method, a 2D Markov chain is adopted to generate an efficient solution to linearize mixed-integer nonlinear programming problems about average task completion time. The other track concentrates on methodologies for big data storage. Dou et al. proposed a privacy-aware cross-cloud service composition method [8]. In the method, evaluations of services are promoted by quality of service (QoS) history records to enhance credibility. Also, a k-means algorithm is adopted as the filter for representative historical records. Yin et al. proposed an efficient storage model based on code erasing [9]. In the proposed model, the large amount of data is separated into various storage nodes, and different users can set different coding parameters to improve the robustness of the overall storage system. Yang et al. proposed a novel approach for error detection in big sensor data [1]. The approach first classifies typical data errors. Then clustered wireless sensor networking (WSN) is introduced for fast error detection based on scale-free network topology. Most detection operations are carried out in spatial data blocks, which ensure its timeliness. To sum up, even though much research in these two directions has been carried out, there are few comprehensive designs to solve problems in ambient assisted living (AAL). The necessity of big data analysis was already pointed out by Mao et al. [11]. A universal method for big data abstraction was also presented in the article. Forkan et al. also proposed a context-aware monitoring based on big data techniques [12]. However, most traditional data processing algorithms use firstin first-out (FIFO) queues to carry data [13]. Such a queue is effective with powerful processors, but it may be that a processor cannot meet the required speed in some scenarios. Under this circumstance, massive data are left in the queue, and cannot be processed quickly, so essentially a breakthrough is necessary to solve the problem. Actually, we have delved into the research of data processing in ehealth networks before, and proposed a local data processing architecture in [14]. In that article, we also adopted the RVNS algorithm as a methodology for data processing, but we explore more essential characteristics about RVNS now, and propose the concept of an RVNS queue. In addition, we consider more about big data transmission in mobile ehealth networks, and propose a interests-matching mechanism to ensure the reliability of data transmission. Based on the above description, an interests-matching RVNS queue architecture is introduced in this article to help health care providers identify valuable data more efficiently. The rest of the article is organized as follows. The main idea of RVNS is presented. We demonstrate how IRQA ensures reliable rapid processing through a three-layer architecture with simulation results. We present a conclusion. Overview of RVNS Algorithm In this section, we introduce RVNS and its original algorithm, Variable Neighborhood Search (VNS). VNS is a meta-heuristic algorithm for solving combinational optimization problems [3]. Normally, for a combinational optimization problem, there is always a solution space S = {x 1, x 2, x 3,, x n }, and an optimal solution selected from S based on some specific utility function. However, in VNS, there are two kinds of optimal solutions, called global and local optimal solutions, denoted as x and x, respectively. A global optimal solution is the currently found x S contributing to the maximum result of x. For every global optimal solution x, there is a neighborhood structure N(x) constructed based on x. N(x) is a subset of S. Also, the neighborhood structure consists of k neighborhoods. These specific IEEE Network January/February

3 Solution space Global optimal solution: X Neighborhoods Figure 2. Overall idea of RVNS. Randomly generate X < X neighborhoods are denoted as N k, where k [1, k max ]. Particularly, when x is a feasible global optimal solution, these k specific neighborhoods are denoted as N k (x). N k is reconstructed continuously through a series of predefined rules when a new feasible global optimal solution is found. A new feasible global optimal solution is always generated from local optimal solutions. In VNS, if we cannot find another x i in a specific neighborhood satisfying (x ) < (x i ) using a specific subroutine, x will be regarded as a local optimal solution of those neighborhoods, where i [1, n]. However, RVNS simplifies the adoption of a subroutine, and just takes an element in current neighborhoods randomly as the local optimal solution. Considering the necessity of rapid processing, we use RVNS for guidance in reorganizing received data in this article. After the confirmation of the local optimal solution, the original global optimal solution x will be replaced by local optimal solution x if (x) < (x ), so a new feasible global optimal solution is generated. At the beginning of RVNS, a global optimal solution x is generated randomly, and a corresponding neighborhood structure N k (x) is constructed. The first local optimal solution x is generated from N 1 (x). If the original global optimal solution x is replaced by the current local optimal solution x, a new neighborhood structure is constructed based on the new feasible global optimal solution. The second local optimal solution isfound in N 2 (x), and the above work flow is repeated. The procedure of RVNS is presented in Fig. 2. To sum up, the whole execution process of VNS mainly includes two parts: constructing changeable neighborhoods systematically and searching for local optimal solutions. The merits of this algorithm are to reduce the calculation complexity through local search and to avoid local optima through changing neighborhood structure systematically. In this article, we rearrange the data process sequence based on the neighborhood structure in RVNS, which differs from the traditional processing pattern and provides the possibility of real-time processing in the big data background. Interests-Based RVNS Queue Architecture In this section, we introduce how IRQA works in detail. As shown in Fig. 3, IRQA can be divided into three layers: data gathering layer (DGL), data integrating layer (DIL), and data analyzing layer (DAL). DGL takes charge of collecting and storing relevant collected data. Also, it checks the status of mobile nodes, and reports interruptive nodes with a fault-tolerant mechanism. Converged data enter DIL, which is used to Local optimal solution: X X > X check the effectiveness of data. Due to the influence of working status and environment, a learning machine is adopted to optimize the checking process. Also, data is divided into different levels in DIL. In the end, DAL puts data into an RVNS queue and reports valuable data. Data Gathering Layer (DGL) In the proposed architecture, data collected by deployed mobile nodes will be converged to a local server and wait for processing. The server interacts with people directly. Usually, in mobile big data networks, nodes are portable, and they can be carried by people, by transportation, or moving themselves. For energy, battery power is used to for monitoring and moving. The continuous moving of these nodes serves as the basis of interests matching. Each node is assigned a unique index during initialization as its identification in system. After data collection, nodes package the data into packets using node indexes for identification of packets. On the other hand, a local server can retrieve the source node of a packet through looking up the node index on that packet. In addition, considering the difference between sending sequence and receiving sequence in the real world, a timestamp is given before sending a packet. Based on this timestamp, a server can sort the collected data in time sequence. Initially, nodes relay packets according to the original routing algorithm. However, due to the unreliable network condition, some packets have to face network delay. Meanwhile, because of interruptive nodes, a server cannot achieve complete data. Accordingly, an effective fault-tolerant mechanism is necessary. In this article, a new fault-tolerant mechanism is designed using the thought of interests matching, as shown in Fig. 4. Lost packets can be found by referring the node indexes when sorting packets based on timestamp. Then an index interest is generated and broadcast for each lost packet according to the found timestamp and node index. Nodes search relevant packets among their buffer once interests are received. If those packets have not been found, nodes take no action except to help broadcast. On the contrary, the node resends corresponding packets matched with interests. The reason that the interests matching mechanism works is the existence of movable nodes. Initially, the routing algorithm sets up a routing path for packet transmission, but the server does not receive the packets because of a routing break. Then an interest is generated and broadcast. Note that there is an interval between the original relay of a lost packet and receiving interest. During this interval, the routing path changes because nodes are moving continuously. Accordingly, packet retransmission may more possibly reach the server. Data Integrating Layer (DIL) Filter: Even though DIL ensures the completeness of a packet, the correctness of data has not been checked. In the real world, it is hard to keep all the data unchanged due to the influence of the network environment. Accordingly, a filter is designed in this section. To give a reference, a relevant data set is introduced. This relevant data set stores massive historical data collected by various nodes. When packets enter DIL, they are first decapsulated. At this moment, a filter checks representative data according to the range retrieved from a relevant data set, based on which some invalid data can be found. Also, a learning machine is adopted in the filter to accumulate common errors so that similar errors can be detected faster. Corresponding error packets are returned to DGL with 38 IEEE Network January/February 216

4 node indexes and timestamp, where relevant interests are generated and data are recollected. Level Division: The mechanisms above ensure data completeness and accuracy. Then the data with various timestamps from different nodes can be organized as a matrix, which can be regarded as a solution space S of RVNS. Data inside the matrix are put in an RVNS queue later. In the matrix, the column and row represent data collected by different nodes at the same time, and data collected by one node at different times, respectively. Afterward, data in S will be divided into several levels randomly. Note that we do not design a complex algorithm for level division. That is because in a big data scenario, data volume could be unthinkable, and has already increased the processing time. If some algorithms with higher complexity are adopted, the processing time further increases with algorithm complexity. At that time, it is impossible to control the processing time. Also, the size of each level tends to be equal when data volume becomes extremely large based on the probabilistic rule. As a result, all the data in S are divided into corresponding levels, which serve as the foundation of neighborhood construction, or what we call RVNS queue construction. Data Analyzing Layer In this section, we first introduce the idea of risk function. Then we analyze how an RVNS queue is used for rapid data processing as well as its detailed design. RVNS Queue Construction: After level division, a neighborhood structure is constructed according to the current global optimal solution. Specifically, an initial global optimal solution (current global optimal solution) is selected randomly from s, that is, one column from the matrix, denoted as x. Neighborhood structure is constructed according to the level of x, denoted as N(X). Suppose X L(opt) (opt [1, max]), and the level of another column from the matrix is denoted as L(temp) (temp [1, max]); then that column will belong to Data analyzing layer Data integrating layer Data gathering layer N k (X), where k = L(temp) L(opt). When all the data is set in its corresponding neighborhood, original RVNS queue construction is finished. Unlike other queues, an RVNS queue is modified during the running of RVNS, which is called neighborhood reconstruction. Before the start of RVNS, every column is assigned into a specific neighborhood according to X. At this moment, data in the solution space are not carried by a traditional FIFO queue anymore, but in the RVNS queue. Since every application scenario analyzing rule has a given threshold, the initial global optimal solution is processed by adopting the corresponding rule, and the result of the initial global optimal solution is recorded. When a new RVNS turn is launched, a function is called to generate a local optimal solution X from the first neighborhood randomly. If the result of X is larger than that of x after being processed by the rule, the relevant data of X is more valuable. In this case, current global optimal solution x will be replaced by local optimal solution X, and the RVNS queue is reconstructed according to the new global optimal solution just as for neighborhood reconstruction. Specifically, Figure 3. Structure of IRQA. Processing unit Valuable data Buffer 1:9 Action A1 1:9 Action A2 1:9 Action A3 1:9 Action A4 1:9 Action A 1:9 Actio Filter Packet Sensor index Timestamp Health care provider RVNS queue Interests Levels when the result is large enough, its relevant data is reported. On the other hand, if the result of current local optimal solution X is smaller than the current global optimal solution, RVNS continues to search in the next neighborhood. Detailed Design of an RVNS Queue: There are two factors that can influence the performance of an RVNS queue. The first factor is the size of the solution space. Because RVNS randomly picks data in the RVNS queue to process, the possibility of selection risk decreases with the expansion of the solution size. On the contrary, if the solution space is too small, the RVNS queue is essentially close to a traditional FIFO queue. Another factor is the interval of neighborhood reconstruction. Since newly generated data enter DAL continuously, the newest data cannot be processed promptly if the interval is too large. On the other hand, data in the solution space do not have enough processing time if the interval is too small. In addition, cross impact may exist between these two factors. For example, assuming that the solution space is relatively small, the time needed for analysis would be definitely shorter than the time needed for analysis with a larger solution IEEE Network January/February

5 A buffer Data packet A5 Source: A Timestamp: 5 A buffer Data packet A5 Source: A Timestamp: 5 Blocked Figure 4. Fault-tolerant mechanism: a) original transmission process; b) fault tolerant transmission process Delivery ratio (%) C C A A Interruptive node space. Thus, without considering the current size of the solution space, a proper interval of neighborhood reconstruction cannot be achieved. Based on this, a group of simulations is designed to demonstrate their optimal value. In the end, we need to pay attention to the terminal condition of RVNS. There are two conventional terminal conditions: maximum time of neighborhood reconstruction and maximum CPU running time. If we adopt the former as the terminal condition, RVNS will stop once the maximum time is achieved. At this moment, because the newest data still need to be analyzed, the former larger results and their relevant data are invalid, which increases the amount of neighborhood (a) (b) D D Blocked B B E Interruptive node Figure 5. Delivery ratio of original scenario and interestsbased scenario. E Original Interests-based Number of faulted trunk nodes reconstruction. On the other hand, if the maximum CPU running time is adopted as the terminal condition, this problem is still unsolved. In this case, it is notable that once a result is larger than a given threshold, IRQA sends a report, and the result must be at its peak at this time. However, it cannot ensure that there is no result larger than the threshold and smaller than the current highest risk factor, which means other smaller results are ignored in the following analysis. In this case, relevant data are ignored without a report. Accordingly, we need to change the conventional terminal conditions, and as described above, the terminal condition should be set as the sending of a report, because the report signal is still generated in a new RVNS queue with a result larger than the threshold and smaller than the former peak. Performance Evaluation The performance evaluation is divided into two parts. The first part evaluates the effectiveness of the interests-based mechanism by virtually setting up a network environment. The second part evaluates the performance of an RVNS queue. Interests-Based Mechanism Evaluation: In this section, we test the improvement resulting from the interests-based mechanism by observing the delivery ratio, which is the ratio of received packets to total generated packets. We first set up a network using C++. In the simulated network, there are 1 movable nodes, each of which represents a set of sensors carried by an elderly person. These nodes move in a 1 m 1 m area with speed varying from 3 m/s to 5 m/s. To assist the transmission, we also set 1 fixed nodes as the trunk nodes. The communication radius of both the moveable nodes and the trunk nodes is 7 m. To simulate the interruption, we gradually reduce the number of available trunk nodes from 1 to 5, and compare the delivery ratio of the original scenario and interests-based scenario in 5, s. The results are presented in Fig. 5. As we can see, the interests-based mechanism can obviously improve the delivery ratio. The reason for the improvement is that when a node carrying lost packets receives corresponding interests, the lost packets are resent. In this way, the environment of the node is favorable after moving for a while. RVNS Queue Evaluation: The performance of an RVNS queue is evaluated through simulation in this section. In this article, a server is simulated using C++. Some of the simulation data come from the Arrhythmia Data Set in the UCI Machine Learning Repository [15]. In simulation, for every 5 s, the server receives a group of packets, the number of which reaches the factorial of 2 (2!). Each group of data enters an RVNS queue with four levels and a FIFO queue, respectively. Data is processed within an assumed rule with complexity of O(2 n ), and the processing time of each data group varies from 5 s to 12 s (integer). The maximum analyzing result is 1, and the given threshold is 9. Simulation length is 5, s. We designed three groups of simulation. The first group finds the optimal parameters of an RVNS queue by changing the size of the solution space and the interval of neighborhood reconstruction, and the results are shown in Fig. 6a. The second group compares the increment of maximum results with time between an RVNS queue and a FIFO queue, and the results are shown in Fig. 6b. The third group compares the waiting time between an RVNS queue and a FIFO queue when average data processing time changes, and the results 4 IEEE Network January/February 216

6 are shown in Fig. 6c. Specifically, in the third group, the average processing time changes from 5 s to 15 s (integer). Based on the analysis above, due to the existence of cross impact, two parameters change together in Fig. 6a. Results show that the optimums of reconstruction interval and solution space size are 1 s and 1, respectively. Thus, these values are used in the following simulation. In addition, Figs. 6b and 6c reveal that the RVNS queue is stable enough to generate quick response when processing time increases with the amount of data. The most important reason is that the sequence of data processing changes because of the RVNS queue so that the potential valuable data can be analyzed and reported early. Conclusion In the big data era, ubiquitous sensing is not always a hard problem. However, the real problems are how to transfer collected data and analyze them efficiently. In this article, a three-layer architecture is proposed for a mobile ehealth network, where a fault-tolerant mechanism is proposed using interests matching to ensure the reliability of network transmission. In addition, filtered data are set in an RVNS queue for rapid processing in the data analyzing layer, and only valuable data are returned to health care providers, which saves people effort in identifying them. In further research, we will focus more on the specific design of the fault-tolerant mechanism to meet the needs of more universal networks. Acknowledgments This work is supported by NSFC ( , , ); SFDPH ( ); NSF of Jiangsu Province (BK ), NUPT (NY21497, XJKY1411); Open Research Fund of Key Lab of Broadband Wireless Communication and Sensor Network Technology (Nanjing University of Posts and Telecommunications), Ministry of Education (NYKL2157); Educational Commission of Guangdong Province (213KJCX131); Guangdong High-Tech Development Fund (213B14135); the 214 Guangdong Province Outstanding Young Professor Project and 213 Top Level Talents Project in the Sailing Plan of Guangdong Province; projects 2479F/2 funded by the Research Council of Norway, and the European Commission FP7 Project CROWN (grant no. PIRSES-GA ). References [1] C. Jardak, P. Mahon, and J. Riihijar, Spatial Big Data and Wireless Networks: Experiences, Applications, and Research Challenges, IEEE Network, vol. 28, no. 4, 214, pp [2] A. Hristova, A. M. Bernardos, and J. R. Casar, Context-Aware Services for Ambient Assisted Living: A Case-Study, 1st Int l. Symp. Applied Sciences on Biomedical and Commun. Technologies, 28, pp [3] P. Hansen and N. Mladenovic, Variable Neighborhood Search: Principles and Applications, Euro. J. Oper. Res., vol. 13, no. 3, 21, pp [4] Y. Zhang et al., Cap: Community Activity Prediction Based on Big Data Analysis, IEEE Network, vol. 28, no. 4, 214, pp [5] L. Kuang et al., A tensor-based Approach for Big Data Representation and Dimensionality Reduction, IEEE Trans. Emerg. Topics Comp., vol. 2, no. 3, 214, pp [6] J. Yu, F. Jiang, and T. Zhu, RTIC-C: A Big Data System for Massive Traffic Information Mining, IEEE Int l. Conf. Cloud Computing and Big Data, 213, pp [7] L. Gu et al., Cost Minimization for Big Data Processing in Geo-Distributed Data Centers, IEEE Trans. Emerg. Topics Comp., vol. 2, no. 3, 214, pp [8] W. Dou et al., Hiresome-ii: Towards Privacy-Aware Cross-Cloud Service Composition for Big Data Applications, IEEE Trans. Parallel Distrib. Sys., vol. 26, no. 2, 213, pp [9] C. Yin et al., Robot: An Efficient Model for Big Data Storage Systems Based on Erasure Coding, 213 IEEE Int l. Conf. Big Data, 213, pp [1] C. Yang et al., A Time Efficient Approach for Detecting Errors in Big Sensor Data on Cloud, IEEE Trans. Parallel Distrib. Sys., vol. 26, no. 2, 215, pp Processing time (s) Result values Waiting time (s) Figure 6. Simulation results: a) processing time with different reconstruction interval and solution space size; b) change of highest founded result with time; and c) waiting time of data when average data processing time changes. [11] R. Mao et al., Overcoming the Challenge of Variety: Big Data Abstraction, the Next Evolution of Data Management for AAL Communication Systems, IEEE Commun. Mag., vol. 53, no. 1, Jan. 215, pp [12] A. R. M. Forkan et al., BDCaM: Big Data for Context-aware Monitoring A Personalized Knowledge Discovery Framework for Assisted Healthcare, IEEE Trans. Cloud Comp., vol. PP, no. 99, 215, pp [13] S. Prakash, H. L. Yann, and T. Johnson, A Nonblocking Algorithm for Shared Queues Using Compare-and-Swap, IEEE Trans. Comp., vol. 43, no. 5, 1994, pp [14] K. Wang et al., LDPA: A Local Data Processing Architecture in Ambient Assisted Living Communications, IEEE Commun. Mag., vol. 53, no. 1, Jan. 215, pp [15] H. A. Guvenir, B. Acar and H. Muderrisoglu, Arrhythmia Data Set in UCI Machine Learning Repository, UC Irvine, 1998; uci.edu/ml/datasets/arrhythmia. Biographies Reconstruction interval (s) (a) Solution space size Time (s) (b) Kun Wang [M] is an associate professor in the School of Internet of Things, Nanjing University Posts and Telecommunications, China. He received his Ph.D. degree from the School of Computers, Nanjing University of Posts and Telecommunications in 29 and was a postdoctoral fellow in the Electrical Engineering Department, University of California, Los Angeles (UCLA) from 213 to 214. He has published over 5 papers in related international conferences and journals, including IEEE Transactions on Industrial Informat RVNS FIFO 2 4 RVNS FIFO Average processing time (s) (c) 1 IEEE Network January/February

7 ics, the IEEE Sensors Journal, IEEE Communications Magazine, IEEE GLOBE- COM, and IEEE ICC. His current research interests include wireless sensor networks, delay-tolerant metworks, stream computing, ubiquitous computing, mobile cloud computing, and information security technologies. He is a member of ACM. Yun Shao is a postgraduate student in the Viterbi School of Engineering, Department of Computer Science, University of Southern California (USC). He received his Bachelor s degree at Nanjing University of Posts and Telecommunications. His current research interests include wireless sensor networks, delay-tolerant networks, and stream computing. Lei Shu [M] (lei.shu@live.ie) received his Ph.D. degree from the Digital Enterprise Research Institute, National University of Ireland, Galway, in 21. Until March 212, he was a specially assigned researcher in the Department of Multimedia Engineering, Graduate School of Information Science and Technology, Osaka University, Japan. In October 212, he joined Guangdong University of Petrochemical Technology, China, as a full professor. In 213, he started to serve at Dalian University of Technology as a Ph.D supervisor. Meanwhile, he is also working as vice director of the Guangdong Provincial Key Laboratory of Petrochemical Equipment Fault Diagnosis, China. He is the founder of the Industrial Security and Wireless Sensor Networks Lab. His research interests include wireless sensor networks, multimedia communication, middleware, and security. He was awarded the IEEE GLOBECOM 21 and ICC 213 Best Paper Awards. He has been serving as Editor in Chief for IEEE CommSoft E-letter and EAI Endorsed Transactions on Industrial Networks and Intelligent Systems. Chunsheng Zhu received his B.E. degree in network engineering from Dalian University of Technology in June 21 and his M.Sc. degree in computer science from St. Francis Xavier University, Antigonish, Nova Scotia, Canada, in May 212. He has been working toward his Ph.D. degree in the Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, Canada, since September 212. He has around 4 papers published or accepted by refereed international journals (e.g., IEEE Transactions on Industrial Electronics, IEEE Systems Journal) and conferences (e.g., IEEE ICC). His current research interests are mainly in the areas of wireless sensor networks and mobile cloud computing. Yan Zhang [SM] is currently head of the Department of Networks at Simula Research Laboratory, Norway, and an associate professor (part-time) at the Department of Informatics, University of Oslo, Norway. He received a Ph.D. degree from the School of Electrical and Electronics Engineering, Nanyang Technological University, Singapore. He is an Associate Editor or on the Editorial Boards of a number of well-established scientific international journals, such as Wiley Wireless Communications and Mobile Computing. He has also served as a Guest Editor for IEEE Transactions on Smart Grid, IEEE Transactions on Industrial Informatics, IEEE Communications Magazine, IEEE Wireless Communications, and IEEE Transactions on Dependable and Secure Computing. He has served or is serving as Chair of a number of conferences, including IEEE PIMRC 216, IEEE CCNC 216, WICON 216, IEEE SmartGridComm 215, and IEEE CloudCom 215. He has served as a TPC member for numerous international conference including IEEE INFOCOM, IEEE ICC, IEEE GLOBECOM, and IEEE WCNC. His current research interest include wireless networks and reliable and secure cyber-physical systems (e.g., healthcare, transport, smart grid). He has received seven Best Paper Awards. 42 IEEE Network January/February 216