Search

Four major design factors of HDFS, the block size, the number of data nodes, the number of client processes and replication factor are investigated to find out the effects on the I/O performance of HDFS by performing experiments in a real physical HDFS infrastructure consisting of 64 Hadoop data nodes of Intel i9 based blades. The block size is observed to be optimal when it equals to about 1Gb or 128MB that is the amount of the data the hard disk drive device can effectively input and output for 1 second in most of today’s off-the-shelf computers. Sophisticated allocation strategy is required to determine the number of mappers and reducers as the number of data nodes increase because the overall performance is influenced in complicated manner by the number of raw data blocks of the job to be processed, the processing time of the blocks for each node and the overhead of shuffling. Experiments shows that Hadoop distributes the work properly that the number of clients does not have a significant impact as the number of clients increases. There is little delay in copying the replica because replication is done in pipelined manner although the network is overloaded.

Modern server systems are usually composed in the form of cluster systems in order to serve not only as many users but also as many kinds of applications as possible. The progression of the cluster system architecture leads in a middleware approach based on the Java framework. The middleware approach alleviates the efforts for the construction and the management of a server system but still preserves its performance and applications on the server. In this research, we introduce a new clustering scheme for the easy construction and maintenance of a cluster server system with the Java Management Extensions. We first demonstrate the construction and configuration process. Our experiment sets can verify that it is easy to construct, expand and manage a middleware based cluster system as well as the applications which reside on it. In addition, we can achieve reasonable performance on our service oriented clustered system with the help of state-of-the-art middleware. The experimental results of performance demonstration contain the availability of a server, and the effectiveness of load balancing and scheduling mechanisms. Especially, our service oriented scheduling mechanism was shown to successfully manage load imbalance under the normal load and cope with the overloaded situations, compared with other known scheduling mechanisms.

Network direct attached storage (NDAS) is a network storage architecture that allows direct attachment of existing ATA/ATAPI devices to Ethernet without a separate server. Unlike other architectures such as NAS, SAN, and USB mass storage, no server computer intervenes between the storage and the client hosts. We describe an NDAS disk controller (NDC) amenable to low-cost single-chip implementation that processes a simplified L3/L4 protocol and converts commands between ATA/ATAPI and Ethernet, while the remaining complex tasks are performed by remote hosts. Unlike NAS architectures that use TCP/IP, NDAS uses a TCP-like lean protocol that lends itself well to high-performance hardware realization. Thanks to the simple NDAS architecture and protocol, an NDC implemented on a single 4 mm /spl times/ 4 mm chip in 0.18 /spl mu/m CMOS technology achieves a maximum throughput of 55 Mbytes/s on gigabit Ethernet, which is comparable to that of a high-performance disk locally attached to a host computer.

We present an application of efficient evolutionary algorithm on the time optimal trajectory planning of a manipulator. The proposed method is simple, flexible and provides a near optimal solution among the huge trajectory space. We first outline the differences in the trajectory parameterization method between the evolutionary algorithm and the conventional nonlinear programming based approach. Then, we propose two novel evolutionary schemes and investigate their characteristics. The first scheme is based upon the canonical genetic algorithm (GA) and the penalty function method to tackle the torque constraints. The second scheme, which is basically based on evolutionary strategy (ES), incorporates effective heuristics to further speed up the search and utilizes the time scaling method to deal with the constraints on actuator torques. The effectiveness and validity of the proposed method are demonstrated through the simulation study in comparison with other conventional nonlinear programming based methods.

As the development of digital broadcasting technologies introduced a new computing platform which utilizes the potential of increasing digital multimedia contents, the software platform for this digital broadcasting infrastructure was expected to open up a new eco-system called interactive TV service. However, despite of successful early-stage development and standardization, the existing interactive TV technologies were not able to form a unified, globally interoperable eco-system lacking the critical mass of popularity from the industry. This research and development work analyzes the limitations of existing interactive TV specifications and provides a new system architecture which seamlessly merges the interactive TV functionality with mobile software platform. As a case study of the suggested converged software architecture, the mobile Android system is analyzed and extended with digital broadcasting functionalities to be a next generation interface TV software platform which makes the interactive TV as an interoperable service with mobile smart phone eco-system.

In order to protect network from the Denial of Service attacks which sends excessive traffic to a host, it is required for network components to throttle unauthorized traffics. The attacker must be identified through the cooperation of routers and must be isolated by the nearest router. It is the most important to identify and isolate the attacker since the nearest router can make ideal blocking of the DoS attacks. In this research, we will present a protocol which can identify the attacker of DoS by the request of victim in cooperation of routers on the attacking path between a victim and the attacker. The performance of our protocol will be verified by simulations and the experiments show that it takes considerably small time to identify the location of attacker.

To serve for high volume of requests with steady availability, servers usually incorporate clustering technology for the availability as well as load distribution, scalability, etc. In this paper, we introduce a new scheme for efficient cluster configuration and management using Java Management Extensions.

A new easy-to-understand physical interpretation of complicated robot control system is given in this paper. Theorems are provided to derive the new physical interpretation of the robot arms as passive dynamical system and the controller as hyperstable system. Utilizing the fact that the multilayer feedforward networks are universal approximators to any continuous function, this new view expands its range of application to neural net type controller and possibly, other AI based controller like fuzzy logic controller. >