Published papers

© 2018 The Korean Institute of Communications and Information Sciences (KICS) Neural Network Learning (NNL) is compute-intensive. It often involves a dropout technique which effectively regularizes the network to avoid overfitting. As such, a hardware accelerator for dropout NNL has been proposed; however, the existing method encounters a huge transfer cost between hardware and software. This paper proposes Slightly-Slacked Dropout (SS-Dropout), a novel deterministic dropout technique to address the transfer cost while accelerating the process. Experimental results show that our SS-Dropout technique improves both the usual and dropout NNL accelerator, i.e., 1.55 times speed-up and three order-of-magnitude less transfer cost, respectively.

© 2020 The Institute of Electrical Engineers of Japan. A smart community utilizes information technology to interconnect and manage community infrastructures. Smart community networks should support a large number of Internet of Things (IoT) devices in community infrastructures to provide services such as smart grids and health monitoring systems. In comparison to cloud-based solutions, smart community services can be deployed in the edge computing area to reduce service latency and to encapsulate private and local information. Furthermore, smart community services can leverage network virtualization technologies to support IoT network services at the edge. A service-oriented container-based solution that processes data streams from IoT sensors using conventional hardware will improve the compatibility and latency of these virtualized network services at the edge. To this end, a software-based edge computing node, namely, the smart community edge (SCE), was proposed to develop a platform for smart community services. SCE supports data-tapping applications, especially for IoT devices, and has a stream processing feature with a comparatively shorter processing delay. This tapping and processing function was named multi-service authorized stream content analysis. SCE captures network stream data and enables service applications using shared memory buffers for a shorter processing delay. SCE supports services as Docker containers to provide remote deployment, service compatibility, and service isolation. SCE allows IoT services to run at the edge through conventional hardware devices, thus, reducing the service latency for delay-sensitive services, which approximately require to sustain latency less than 10 ms. The proposed SCE achieves 10 Gbps bandwidth with a 16 core server when compared to the f-stack library with a 5 Gbps bandwidth. SCE deployment on conventional hardware devices shows its capability of operating at 1-10 Gbps line rates to support up to eight services at 500 Mbps data bandwidth per service, while keeping the overall latency below 1 ms. Therefore, SCE provides a platform for delay-sensitive IoT services at the network edge.

© 2019 The Institute of Electrical Engineers of Japan. Fog services have been gaining momentum due to the demand generated by IoT-based services that require data and services to be provided at the edge of the network. This phenomenon creates several problems for Internet systems, such as security, and data management. Particularly, edge and fog computing addresses these problems and provides new services by solving these issues. To this end, a service-oriented router (SoR) was proposed and extended to provide edge and fog services by performing authorized stream content analysis (ASCA). Packet Capture (Libpcap) library was used to develop the initial implementation of the SoR to perform ASCA using conventional hardware and software platforms. However, Libpcap implementation together with conventional hardware demonstrated limited performance. Consequently, this paper proposes a method using Intel technologies such as the Intel Data Plane Development Kit and Hyperscan to accelerate stream processing and string matching on SoRs. The paper compares both implementations of SoRs and evaluates the performance with regard to packet throughput, CPU utilization, and memory usage. Furthermore, the latency of major functions is evaluated under Libpcap-based and DPDK-based SoR implementations. The results demonstrate an increase of more than 0.8 Gbps in bandwidth in a 1 Gbps link using less hardware utilization when SoR is implemented using Intel technologies compared to the original Libpcap implementation.

Kinetic characteristics of thermal energy storage (TES) using tetrabutylammonium acrylate (TBAAc) hydrate were experimentally evaluated for practical use as PCMs. Mechanical agitation or ultrasonic vibration was added to detach the hydrate adhesion on the heat exchanger, which could be a thermal resistance. The effect of the external forces also was evaluated by changing their rotation rate and frequency. When the agitation rate was 600 rpm, the system achieved TES density of 140 MJ/m3 in 2.9 hours. This value is comparable to the ideal performance of ice TES when its solid phase fraction is 45%. UA/V (U: thermal transfer coefficient, A: surface area of the heat ex-change coil, V: volume of the TES medium) is known as an index of the ease of heat transfer in a heat exchanger. UA/V obtained in this study was comparable to that of other common heat exchang-ers, which means the equivalent performance would be available by setting the similar UA/V. In this study, we succeeded in obtaining practical data for heat storage by TBAAc hydrate. The data obtained in this study will be a great help for the practical application of hydrate heat storage in the future.

With the active development of IES in standards since the mid-2010s, the society has made considerable progress with multiple standards' developments. This paper presents the results of engaging in standards' development within and across borders of an IEEE society. In particular, the hands-on INTEROP Plugfests, coupled with the CoEs, provide platforms to create ideas for standards, develop standards, initiate interoperability among multiple vendors, providing competitive time-to-market advantage for involved industry partners.

This paper proposes a service-oriented communication platform enabling the provision of scalable smart community services. The proposed platform is managed using Extensible Markup Language (XML) that has advantages in terms of flexibility, cost-effective implementation and affinity with databases. In practice, the platform is embedded into distributed communication nodes, e.g. routers, switches and home gateways, and both existing IP services and smart community services are provided over public networks. In this research, as an application, a building energy management system (BEMS) using the platform is demonstrated. The feasibility of the network architecture with the platform is discussed based on experimental results of data acquisition and control in the BEMS demonstration. © 2012 IEEE.

The Service-oriented Router (SoR) is a router for next generation networks. This router has a high-throughput database and it has the capability to snoop all transaction in its interfaces. The SoR also provides an API which can provide stored content in order to enrich services. The SoR has the potential to shift the current Internet infrastructure to an information-based, open innovation platform. In order to evaluate the efficiency of SoR-based applications and services, a specific simulator is required. Though the NS2 simulator does not simulate payload handling, it is a well-known network simulator and there are many resources available for evaluations. In this research, we implemented three new modules for the NS2 platform to test the concepts of an SoR, an SoR agent, a packet container and a packet analyzer. The SoR agent can interconnect the application and physical layers of the NS2 module with an additional function that can handle user data as a real payload and attaches it to the IP packet. The packet container is an improved NS2 module that can attach user data and trace information. A successful simulation has been performed for end-to-end packet transfer with a 0.02s packet delivery delay compared to typical NS2 packet delivery. The NS2 router has been enhanced for packet stream analyzing and content centric packet re-routing. Content centric packet re-routing has been achieved with 0.01s additional delay for end-to-end packet delivery. Result shows that Deep Packet Inspection and content centric networking in an SoR-based network can be achieved. © 2013 The Authors. Published by Elsevier B.V.

Internet users are constantly demanding faster and higher quality services from their internet service providers. Therefore, for fast data delivery of such applications, Content Delivery Networks (CDNs) have been introduced. Most CDN providers use Domain Name Resolution (DNS) based request routing (RR) methods to find the nearest server for a particular client and it has both advantages and disadvantages. Importantly, disadvantages result high latencies of data delivery and network congestions. To maintain rich information in the Internet and to shift the current Internet infrastructure to an information-based open environment platform, Service-oriented Routers (SoRs) have been introduced. An SoR has a high-Throughput database and it is able to analyze all transactions on its interfaces. Therefore, we have used the basic functionalities of the SoR to diminish disadvantages of the DNS-based RR methods. Proposed system is independent from DNS-based RR and we have conducted experiments based on content-centric RR using the SoR basic functionalities and successfully evaluated and compared both of the round trip time (RTT) and the packet inter arrival time. Our results indicated that SoR-based method can reduce upto 40-50% latency in both connection initiation time and content migration time in-between servers. © 2013 IEEE.

© 2016 IEEE. The growth of energy consumption in data centers is becoming one of the significant problems all over the world. The power usage of IT devices and air-conditioning (AC) units accounts for up to three-quarters of total energy consumption in typical data centers. This study discusses the effectiveness of using shutter-controlled air flow management to control cold air flow into the servers. It can impact the power consumption of both, the servers and the AC units. A raised floor data center has two types of aisles - cold and hot. In most cases, excessive volumes of cold air are pumped into the servers using cold aisles. We are proposing the utilization of a shutter to control the air flow and reduce wastage while preventing damage due to increased CPU temperature. The shutter is intended to control the air flow and maintain the temperature appropriately according to the workload of the servers. We evaluated the proposed system for controlling the flow of air into the servers. The results show that cold air volumes can be reduced by up to 92% while maintaining appropriate control of CPU temperature.

© 2015 ACM. Packet propagation delay reduction is becoming the primary concern of the Internet. Internet service providers (ISPs) attempt to optimize packet routing to offer the best route to their subscribers by achieving desired network performance. Consequently, a route computation metric that uses packet propagation delay instead of link state will be a strong incentive for ISP routing optimization, in which case ISPs could ensure minimal delay route paths for their subscribers. To this end, we present SLRouting, a novel Interior Gateway Routing Protocol. SLRouting calculates a composite route metric using packet waiting delays of servers and routers as well as the packet propagation delay of network links. SLRouting computes the route matrix by selecting the minimal delay path for destination networks. This paper presents the first version of the SLRouting including its theory, design, and implementation notes. A prototype of the proposed protocol is implemented using the ns-3 simulator, and the results were used to evaluate the proposed protocol.

We propose a sophisticated synthesis methodology for SoC (System-on-Chip) architectures from the system level specification based on reusable high-level IPs named as Virtual Cores (VCores), in this paper. This synthesis methodology generates an initial architecture that consists of a CPU, buses, IPs, peripherals, I/Os and an RTOS (Real Time Operating System), as well as making tradeoffs to the architecture, between hardware and software on assigned software VCores and hardware VCores. The results of an architecture level design experiment, using the proposed methodology, shows that the partial automation of the architecture synthesis process, allied with design reuse, accelerates the architecture design, therefore, reducing the time required to design an architecture of SoC.

Internet users are constantly demanding faster and higher quality services from their internet service providers. This results an increase in broadband services, and access to delivery of applications enriched in data become top priority. Therefore, for fast delivery of such applications, Content Delivery Networks (CDNs) have been introduced. The Internet has become a massive system consisting of enormous amounts of information. To maintain the rich information in the Internet and to achieve maximum benefit from networks, Service-oriented Routers (SoRs) have been introduced. A SoR has a high-throughput database and is able to analyze all transactions on its interfaces. In addition, SoRs can provide APIs for accessing stored contents in order to enrich services. CDNs generally use Request Routing (RR) methods to achieve low user latency by performing load balancing among servers. RR is accomplished by selecting the nearest server for a particular user. In this paper we implement a new RR infrastructure that is based on the content of packet streams. The proposed system is independent of DNS-based RR, which is the predominant RR method used by current CDN infrastructures. In our system, content-centric request re-routing with SoR is used. Experimental evaluation and comparison of the round trip time (RTT) result of our proposed system with DNS-based RR indicates that our system achieves 50-60% reduction in connection initiation time. The SoR based RR method is able to reroute packets without disturbing clients, which is not possible using existing RR methods. This feature will make future CDNs more effective and convenient. © 2012 IEEE.

The Future Internet architecture would be advanced and flexible depending on service requirements by program. The concept of Software-Defined Network (SDN) has a possibility to enable innovation while hiding much of the complexity on the networking design. SDN is mainly used by OpenFlow in real scenario, which can use up to layer-4. In this paper, we propose SDN which consider up to layer-7. To maintain flexible network and to get maximum benefit from networks, Service-oriented Router (SoR) was introduced. A SoR enables layer-2 to layer-7 information to be captured, analyzed, and stored and has a high-throughput database(DB) and is able to analyze all transactions on its interfaces. In addition, SoRs can provide APIs to access stored contents in order to enrich services. In our system, we make network programmable by considering layer-7 information, which can not be captured by using OpenFlow method. This feature will make future SDN more effective and convenient. © 2013 IEEE.

© 2016 IEEE. For the development of a smart grid, smart meter is a key device to measure the electric power usage of network-connected houses. Smart meters are currently being installed into households, and play the important role for providing a demand response service. A demand response is a necessary service in order to adjust supply-demand balancing because the balance is kept by the cost in the electricity market. Therefore, the customers of the electricity supply companies are expected to be optimally assigned as a group. In this paper, we separate a set of households into clusters as optimally assigned customers. When conducting the clustering, the utilization of unsupervised learning using data from a smart meter is required. In this study, we propose a method of household clustering for a demand response event by sparse coding, which is a type of neural network. The proposed method generates a power consumption model of each household, finds simple relationship distances between households, and conducts hierarchical clustering based on these distances. In addition, to extract the characteristics of fluctuating load usage, we conduct data normalization that cuts off at a fixed load usage of each household. To confirm the effect of the proposed clustering method, the usage tendency of household air conditioning (A/C) units was evaluated.

Structural Health Monitoring (SHM) system uses a number of sensors for a precise prediction. This prediction enables to provide early indication of vulnerability of a structure and the timing of reinforcement in order to prevent future collapse. SHM systems are now widely spreading in infrastructures. However, conventional SHM systems did not take account of accidental collapse. There is a possibility that sensor node goes down or network goes offline. If sensor nodes are broken by damage, important data, stored in the nodes will have lost. To prevent the data loss, this paper proposes data-distributed fault-tolerant sensor network system for a SHM system. Proposed system consists of three functions, PING (Packet InterNet Groper) based network monitoring function (PNM), backup node selecting function (BNS) and socket communication function. Proposed system is implemented on PCs and practically evaluated in a laboratory environment on terminals of KNIVES, which is a test bed of sensor network system. As a result, Data sharing process on PC and KNIVES was confirmed and proposed method can provide a light-weight fault-tolerant system taking account of accidental collapse. ©2008 IEEE.

© 2003 IEEE. In this paper, we propose a novel architecture synthesis method for SoCs using VCores (virtual cores). VCores are reusable and configurable high-level descriptions. An initial SoC architecture, which consists of a CPU, buses, and peripherals, is generated, based on an architecture template. The hardware and software tradeoff is possible on the architecture model after assignment of software VCores or hardware VCores. The assignment is based on the results of the architecture's performance estimation. We present a prototype of the synthesis for SoC architecture using VCores and an architecture level design experiment using this prototype.

Internet based demand and supply control system is expected as an effective solution for saving fossil fuels by controlling power consumption automatically with taking account of environmental information. KNIVES is one of this systems and it uses distributed and cooperative power control algorithm. In 2007, the latest prototype of KNIVES has been developed and be operated so as to evaluate in real environment. KNIVES aims to improve the total efficiency of energy utilization in general household. The distributed and cooperative control is considered to improve the advantages of demand control system even if the controlled building is a small-scale customer. In this paper, the table-based scheduling method adapted to the distribute DSM is introduced. Furthermore, the enhancement of this method in order to aware environmental amenity is also indicated. © 2008 IEEE.

© 2017 ASME. With the growth of Internet-based services, data centers must deal with increasing data traffic. Load-balancing technologies can help data centers process data effectively and stably; however, the current load-balancing methods do not take into account the heat generated by servers. Excessive heat can increase the failure rate of IT devices and the energy consumption of air conditioning systems, both of which lead to higher data center maintenance costs. This paper aims to simultaneously increase the coefficient of performance (COP) of the data center's air-conditioning equipment and decrease the semiconductor-based equipment failure rate. To do so-and, consequently, reduce the operation and maintenance costs-we propose a novel request distribution system based on servertemperature and evaluate the proposed system by creating a thermal model of a data center. As a result, it is suggested that using the proposed load-balancing method the semiconductor failure rate can be reduced by 32 % when compared with the common round-robin distribution method, and by 19 % when compared with a load-balancing method based on CPU utilization. Moreover, the COP of the air-conditioning equipment obtained with the proposed method is recognized to be higher than those obtained with either the round-robin or the CPUutilization- based methods.