Research On Grid Architechture And Corresponding Key Technologies Based On P2P Overlay

The representatives of Grid systems are Globus and Web services. However, these systems are based on client/server (CS) mode. As the services are published and discovered in centralized mode, they have poor scalability and a single point of failure. Thus, many service-oriented Grid architectures have been proposed, which absorb the advantages of pure P2P and C/S modes. We can see that most of the current research focuses on either P2P based algorithms for transferring data to reduce the burden of a centralized server or utilizes P2P routing algorithms to locate available resources. Thus, they just adapt the existing P2P as an improvement to pure Grid systems, and only introduce it into the resource management aspect of Grid systems. However, the fundamental differences between the different types of P2P systems with regard to the implementation of the information reflection and collection mechanism indicate that not all P2P systems are compatible with a Grid application. There are two types of P2P lookup services widely used for decentralized P2P systems: structured searching mechanism and unstructured searching mechanism. Structured systems implement a distributed index or a distributed hash table (DHT) by using an efficient object-to-node mapping functionality with a hash function. Since the location of objects is based on the exact knowledge of the corresponding name (or key), DHT-like algorithms lack flexibility and fault-tolerance, and are unsuitable for Grid application because users are widely distributed and come from non-cooperating organizations with highly dynamic joining and leaving. To unstructured P2P networks, the large volume of query traffic generated by the flooding of messages limits the scalability and efficiency of this approach. Thus, Supernode-structures which use a hierarchical approach are presented to combine the merits of centralized algorithms (e.g. high efficiency) with those of distributed algorithms (e.g. autonomous and adaptive for dynamic environments). However, the key problem is that most hierarchical structures use tree-based algorithms. Due to the single-parent nature, the tree-based overlay is unbalanced and vulnerable to high "churn" (departure) rates of peers, since the departure of one peer affects all of its children.In our paper, we analyze service-oriented architecture (SOA), P2P, and Grid in detail, and integrate them together to build a fault tolerant architecture. The main work and contributions are presented in the following aspects:(1) The new architecture we propose in this paper is called Information Pool Based Grid Architecture (IPBGA) in which virtual organization is separated into an information service layer and a cooperation layer via a service-oriented method. This information service layer is responsible for collecting information about all the available resources and tasks by information pool protocol (IPP), and then builds a complete view for shared dynamic resources. The structure of the information service layer is in P2P mode. Through local optimization of autonomous system, nodes can cooperate with each other to transfer tasks in the cooperation services layer based on the complete view provided by the information service layer. Therefore, our architecture itself is a hybrid P2P based Grid mode.(2) We present a self-adaptive, self-organizing and self-optimization -so called self -* properties unstructured P2P overlay. The main idea is to combined the merits of DHT P2P overlay (i.e. structured and efficient node structure and query mechanism) with that of an unstructured P2P overlay (i.e. scalability, adaptiveness and robustness) to build a whole new overlay. The key characteristics of our design are summarized below: (a) The unstructured overlay could efficiently guarantee the unique and successive of peers' identities without the need of global knowledge, even though both dynamic and random events like joins, leaves and crashes take place. This is very important for the overlay could be maintained effectively and autonomously so that the corresponding mechanisms (e.g. game theory) could be applied conveniently. (b) A three information center (Tri-IC) mechanism is presented; each IC has a different duty and backup the other in order to lighten the loads of cluster heads in a traditional hierarchical overlay and thus preventing a single fault. It also brings higher search efficiency and is more robust than the popular redundant mechanism. (c) The paper proposes a whole new hierarchy structure, where update and backup only take place in two neighboring levels or layers. The maintenance cost is alleviated because the update and processing costs are lower. It differs from tree-based algorithms in which there are parents and grandpa nodes so that the update and backup information have to send to more than two levels. Small world characteristic which is usually utilized in structured DHT-like overlay is applied to build the unstructured overlay making the response mechanism in the overlay being typically the best way to answer the queries, and provides anonymity and censorship protection with efficiency. (d) Better than clustering algorithms only topology-aware, a simpler location-aware mechanism combined with an information exchange frequency is presented to penalize selfish nodes as free riders. Evaluation proved that MPO is robust, highly efficient and of a low-cost.(3) A new searching algorithm based on the overlay has been presented. It is called BiDirection Search (BDS) algorithm. By a mass of simulations for effects of topologies on searching performance, it is farther improved that the MPO is robust, highly efficient and it is very suitable for many searching algorithms. Based on those simulations' results, BDS is presented. The evaluations show that BDS perform better on the MPO than other representative unstructured searching algorithms.(4) A game theory Hotelling analogy mode is presented to solve non-cooperation behavior, in order to guarantee the validity of incentive and penality mechanisms. Current researches on P2P overlay and corresponding searching mechanisms based on an assumption that all nodes in systems are cooperational, that is to say, one node utilized others' resources or services while it shared its own services. But in fact, there are a lot of non-cooperational nodes such as free riders (A free rider is a node that has access to services and consumes resources without participating in any of them.), cheating behaviors such as whitewashing (peers leave the system and rejoin with new identities to avoid reputational penalties) and malicious attacks such as Sibil Attack (a node creates multiple identities in order to gain control of a part of the system). To solve aboving problems, the thesis presents a Hotelling analogy model based on the research in author's P2P overlay. It can make nodes in the systems to obey the rules including incentive and penalizing mechanisms, so that the efficiency and capability of performance are improved.