Grid Performance Prediction Based On General Regression Neural Network

Benefiting from the concept of Electronic Grid, Grid computing came to birth and developed rapidly in company with internet technique. Along with the huge improvement and rapid progress of computer performance and network communicating technology, the application demands develop towards high- performance, large-scale, diversity and multi-function, thus it becomes a top priority to connect the various and heterogeneous high-performance computing resources, storage, data and other special resources together which are geographically distributed via high speednetwork, to establish high-performance collaborative computing and solve huge application problems cooperatively, which is also called wide-area high-performance meta computing technology, one type of the distributed computing technology. Under the mode of grid computing, Grid infrastructure masks the distinction among hardware platform, operation systems, organizations and regions and gathers various resources into a Virtual Super Computer"which can be shared and managed by different users and organizations transparently and efficiently. Grid computing has powerful data processing capability and is capable of making full use of idle computing resources through internet. As a latest emerging network infrastructure, Grid is the key technology of the new generation of the Internet.The project"The monitoring and prediction of Grid resources and application programs"is aimed to the full utilization of idle resources. The Resource-related problem is the core issue of the Grid Computing. The frequently used resources in Grid are: processing capabilities, storage resources, directories, network resources, distributed file systems, distributed computing pool, computing clusters etc. In Grid Computing environment, resources are distributed among different administrative domain at various locations, possessed and operated by different organization at various deploy and security strategies. The resources are distributed in a very wide geographical area, connected through the WAN, and their huge types and large quantities require a certain amount of cooperative work, but the availability of information and resources is usually dynamic. Under this scenario, the operation of resources includes organization, accession, localization, discovery, scheduling, allocation, authentication, process-creation as well as all the other activities about the preparation of resources. The approach to effectively discovering and inquiring resource information at different administrative domains is the basis part of the monitoring and prediction of Grid resources and the Grid application.The monitoring module is mainly responsible for the real-time, effective monitoring of the resources under the grid systems. This module is mainly implemented with Virtual File System. Virtual File System is a special file system in UNIX. Although it exists in the operating system as a file system, it is in fact a region in the memory, recording the changes of the system parameters on time. It is a safe and efficient method to read the system parameters Through the Virtual File System. It not only guarantees the synchronicity of data, but also avoids the operating of Kernel directly. The main idea of monitoring module is to read the latest information of the system resources within the nodes of the grid system continuously and store the information in the database of one node of the Grid system, and then predict the needed information read by module as its input vector.The predict module is an important part in this project, it use Artificial Neural Network technology. Artificial Neural Network inspired by the theoretical model of the human brain work, simulate the human brain organization to a new information processing system, it is an effective mathematical method to explore multiple factors, complex nonlinear problems. Recent years, with the rapid develop of the artificial neural network technology, Artificial Neural Network Gradually applied to the usual life. In this project, use the system information which is accesses from the grid system to predict grid performance. The author investigates and researches General Regression Neural Network (GRNN) which is a type of Artificial Neural Network, based on the needs of project, the author also improved the traditional GRNN. There are two parts in the new GRNN module: the Radial Basis Function layer and the weight adjustment layer. The first, the sample data through the clustering algorithm to get the offsets and the center, use these values for the first mapping with the function of Radial Basis Function layer, then the second mapping with the function of weight adjustment layer, comparison of the error between the final output through the GRNN module and the real output in the sample data, according to this error adjust neurons'weights and thresholds with back propagation algorithm, the aim is to reduce the predict error. Artificial neural networks applied to the grid computing technology are rare, therefore, system architecture and some ideas in this project are the reference to the grid development.This paper is part of project"The monitoring and prediction of Grid resources and Grid applications"which is funded by National Natural Science Foundation. According to the method described above implements the grid performance prediction system, the performance of prediction System is not bad. A suit of methods and prototype are established, the predict module in this project is implemented with the thinking of code reusability, high cohesion, low coupling, so this module can be used to other predict systems. Cover the short of predict area, it has great realistic significance on how to make good use of the idle resources. The project can be used to direct scheduling system, and rational use of grid resources. In the future, we can do from the following areas to improve and expand the existing prototype system: with the rapid grid develop, Grid users and scheduling system needs many characteristics from grid application, to enhance the accuracy of the forecast information and adaptability, Grid application procedures need to category, set up artificial neural network model for different types of grid application, satisfied the web service users, according to the application characteristics choose the module to predict the application execute time; In the resource monitoring, to improve the current system only supports Linux system; Under the rapid develop of grid technology, how to make good use of the idle resources is very important, next step should be expand the applications of grid performance prediction system.Grid is a new technology, related standard and criterion is being established and improved. The monitoring and prediction of Grid resources and Grid applications is as the cornerstone of Grid infrastructure, Grid information service deserves more applied research and practice.