Catálogo de publicaciones - libros

Grid systems are well-known for its high performance computing or large data storage with inexpensive devices. They can be categorized into two major types: computational grid and data grid. Data grid is used for data intensive applications. In data grids, replication is used to reduce access latency. It can also improve data availability, load balancing and fault tolerance. If there are many replicas, they may have coherence problems while being updated. In this paper, based on the aggressive-copy method, we developed an algorithm using pipeline concept, such that the data transfer tasks can be done simultaneously. This novel Pipeline-based Aggressive Copy method can accelerate the update speed and decrease users’ waiting times. We used Globus toolkit for our framework. Compared with the existing schemes and from the preliminary simulation results, our method shows notable improvement in overall completion time.

Resource sharing computing is a project that realizes high performance computing by utilizing the resources of peers that are connected to the Internet. Resource sharing computing provides a dynamic internet environment where peers can freely participate, but it raises questions on the reliability of operation processing. Existing resource sharing computing stores intermediate operation results in peers’ local disks. Thus, when faults happen on peers’ side, some peers need to wait for processing to reconnect with possibility of considerable delay. In case there is no reconnection, the intermediate operation results cannot be used. In addition, it is difficult to support cooperation due to incompatible modes of operation processing among heterogeneous systems. This thesis is to propose a cooperation management system, and define cooperation and cooperation groups necessary to improve the reliability in the resource sharing computing environment. Cooperation is a series of tasks that involve sorting tasks, processing tasks sequentially, and producing results. Cooperation group is a gathering of peers that can cooperate. Groups are created among different types of systems to enable cooperation among peers within the same group. Also, middle DB Server is proposed in a hierarchical structure to shorten delay and increase the reusability of intermediate operation results. As the intermediate operation results are stored in the middle DB Server in case there occurs a fault on peers’ side, waiting time for reconnection is reduced through cooperation, and the reusability of intermediate operation result is improved. In this paper, we propose a structure that can store intermediate operation result in middle DB Server to improve reliability in resource sharing computing environment, and suggest a design for cooperation group service, discovery service, and task management service of cooperation management system.

We present an index system for locating files or other data objects in a grid environment. The system is constructed using a distributed hash table, and is scalable, fault-tolerant, and self-organizing. The index is dynamically updated to reflect the state of the storage elements, and can hence deal with nomadic data. The system provides extra services to ease integration with other systems: A registrant service integrates existing storage elements into the system, and a query proxy provides an easy way to query the system. A security model, which builds on the existing grid security model is also provided. An implementation has been created and its performance measured. The system is shown to scale as more nodes are added to the system.

In this paper, we have addressed the complex problem of determining a recovery line for cluster federation and have proposed a fast recovery algorithm to handle failures in cluster federations. The main feature of the proposed algorithm is that it can be executed simultaneously by all clusters in the cluster federation. Besides, the number of trips to the stable storage necessary for executing the algorithm is much less compared to the same in some existing works. Also the proposed algorithm does not suffer from any message storm unlike some noted work in this area.

In many distributed applications, each participating node can be characterized by one single set of attributes. The problem is to support complex queries, such as range and k-nearest-neighbor (KNN) queries, on this set of multidimensional attributes. Traditional peer-to-peer (P2P) systems either adopt an unstructured interconnection and use flooding to search for matching nodes, or implement a distributed hash table (DHT) to serve as a directory for indexing the attributes. The former suffers from excessive flooding traffic, while the latter has the overhead of updating and maintaining the directory. This paper introduces an attribute-based P2P interconnection strategy that uses the attributes to interconnect the peers instead of hash keys. Under the condition that each node is characterized by one set of attributes, the attribute-based networks can support range and KNN queries, guarantee lookup efficiency, and eliminate the need to maintain a directory.

It is desirable to replace supercomputers with low cost networks of computers to run computationally intensive programs. To alleviate the burden of writing distributed programs, automatic translation of sequential to distributed programs is highly recommended. In this paper a new architecture to support automatic translation of sequential to distributed programs is offered. A formal specification of the structure and behavior of the architecture components is presented. The applicability of the specified architecture is demonstrated by presenting its implementation details and evaluating the performance of the resultant distributed program.

This article presents a dynamic data distribution method for data storage in a P2P network. In our system named Us (Ubiquitous storage), peers are arranged in groups called Metapeers to deal with account failure correlation. To minimize end user traffic according to the reconstruction process, distribution must take into account a new measure: the maximum disturbance cost of a peer. In a previous work, we defined a static distribution scheme which minimizes this reconstruction cost derivated from affine plan theory. In this paper we extend this distribution scheme to deal with the dynamic behaviour of peer to peer systems.

Today large-scale applications require access to data stored in heterogeneous storage systems located at geographically distributed virtual organizations. In such applications, users are forced to deal with different administrative policies and structures at each site, and various data access mechanisms on each storage system. This implies a lot of human interventions in order to develop dedicated programs and scripts for data transfer between theses systems. This paper presents architecture which enables the automation of data transfers between distributed file systems irrespective of their heterogeneity. This feature enables high-level schedulers integrated with GRAVY to control data placements like computational jobs. GRAVY supports multiple data access protocols and provides an easy-to-use interface for novice Grid users.

One of the challenges of Grid computing is the integration of legacy scientific applications. The Web Services Architecture (WSA) is an ideal technology to integrate legacy applications into the grid environment. Web Services Resource Framework (WSRF) extends Web Services and makes them stateful. Based on WSRF, we implement a framework which utilizes WSRF resource to submit applications and to monitor execution status. We deploy only one Factory Service to create the resources and one Grid Service as the uniform interface for all the applications in each computing resource. We can dynamically deploy some legacy applications in the Grid or remove these applications without stopping the execution of entire system. Moreover, we present an implementation of one meta-scheduler which integrates Grid resources in complex Grid environment.

Many grid architectures have been developed since the first proto-grid systems in the early 70’s, but there are not so many based on groups using an efficient node neighbor selection. This paper proposes a grid architecture based on groups. The architecture organizes logical connections between nodes from different groups of nodes allowing sharing resources, data or computing time between groups. Connections are used to find and share available resources from other groups and they are established based on node’s available capacity. Suitable nodes have higher roles in the architecture and their function is to organize connections based on a node selection process. Nodes’ logical connections topology changes depending on some dynamic parameters. The architecture is scalable and fault-tolerant. We describe the protocol, its management and real measurements. It could be used as an intergrid protocol.