An Architecture For PC Based Grid Computing And Key Technologies

With the rapid development of computer and networking technologies, the cost of owndership of kinds of computing resources (desktop computers, workstations, super computers and clusters), storage resources and network resources have been greatly decreased, which in turn drive the birth and fast development of Grid Computing technology. Today a lot of research interests in Grid Computing have been concentrating on how to aggregate many high-performance computing resources, expensive experimental instruments and hugh storages to construct kinds of super grids that meet the computing needs of scientific research, mathematical modeling and engineering applicationn. In this thesis, we focused on a relatively low-cost approach to realize grid computing, which leverages a lot of personal computers (PC) that are costly available from anywhere, but are lessly used or idle for most of the time. A single PC may not possess fantasic processing power, but if they are aggregated, considerable processing capability can be desired.PC-based grid computing is quite different from high performance grid computing. Normally it is built upon a lot of PCs distributed worldwide, which are voluntarily organized based on specific computing objectives. This is a highly dynamic and autonomous environmentt, which brings considerable challenges in grid architecting, resource management and task scheduling. Now a lot of PC-based grid applications, such as SETI@Home, FightAIDS@Home and GIMPS, follow a simple approach to loosely couple PCs into a virtual organization for solving specific computation problems. But as so far, there is no suitable general-purpose grid architecture and relative platforms or tools to assist the construction of PC grids, which can support various grid applications. There are still a lot of topics to be addressed in order to promote PC-based grid computing.In the thesis, we firstly proposed a three-layered architecture for PC based grid computing, namely PGrid to serve the purpose mentiond above. It contains Agent Control Layer, Communication Infrastructure and Core Service Layer. Agent control layer consists of a lot of agents that reside on PC grid node, which provide a virtual task execution and control environment. Communication infrastructure makes transparent the physical localtions, platform heterogenuity and lower-level communication details of PC grid nodes to applications and provides simple interfacesto interact with them. The core service layer provides grid services to customers, including task submission, assignment, monitoing and result compilation. PGrid simplifies the construction of PC Grid, and addresses its requirements of scalability, robustness and flexibility. Based on PGrid, we proposed serveral key technologies to support its implementation.For communication infrastructure, we proposed a service request broking middleware, namely SRB. It is built on a group of mediation servers that act as an message exchange broker between service requestors and service providers. SRB doesn't care the content of the service request/response message, while just receiving and routing it the corresponding recipients. Based on it, we can flexibly construct different types of grid, such as service or task oriented grids.For agent control layer, we studied the performance prediction models of PC grid node. It can help agent proactively schedule its task assignment based on the heuristics learned from its observation of PC grid node's behavior in the past. We proposed a composite prediction model to address the need of online performance forecast that lacks abundant observation samples, and an approach that combines auto-regressive modeling and neural networks for offline exploration of PC grid node's behavior.We also studied and proposed a multi-agent-based resource negotiation and trading model for agent control layer. There are 2 important intelligent entities in the model: User Agent and Resource Agent, each of which has its own beliefs, desires and intentions. Following different benefit objectives, they coordinate and cooperate with each other in a market environment to settle the deal for resource usage and task assignment. For this model, we also studied the commodity model of PC grid resources that can adjust the prices spontaneously according to the demand-supply equilibrium the market.Finally, we proposed a unified grid-monitoring framework for the core service layer, namely MonitorView, which provides a unified approach to monitor the tasks running on each PC grid node that may come from different applications. MonitorView consists of three parts: unified message interface, application-oriented management protocol and monitoring configuration service. Based on them, MonitorView realizes the flexible management of PC grid regarding the task assignment, execution and system monitoing, which is weak in many PC-based grid computing projects today.