Grid Performance Prediction Based On BP Neural Network

With the advancement of Computer Technology, The demand of Large Scale Computing become more and more while the quality grow higher and higher. In the 1990s, Internet developed very fast, which lead to the promotion of a new Distributed Computing System– Grid. Based on Internet environment, Grid can combine the idle computing resource in the Internet to a virtual"super computer"which is used to deal with Large Scale Computing. That is, Grid is actually a Distributed Computing System set up in the Internet environment. There are two significant problems which might be relative to the performance of the Distributed Computing System: the allocation of the Grid Resource and the Scheduling of the Grid Application. In addition to these two factors, not only the condition of every node should be considered, but also some parameters of Internet. Moreover, Grid allows every node join and quit freely, so that an intensive monitoring to node which brings benefit to rescheduling of the program failure is necessary.Generally, the Grid Resource is consists of the computers in the Internet. So, the Monitoring to the Grid Resource is divided into two parts. One part is the monitoring to the network condition of Internet. In this part, Network Band Width and Network Latency are two important parameters which can describe the Internet condition. The other part is the inside resource usage and load of node which might affect the performance of the node. Thus, these factors should be the factors that should be monitored.The program running in Grid, no matter running in one node or multi-nodes, will be finally allocated to Grid Resources with the form of process. Consequently, the Grid Process Monitoring should contain two parts. That is, the monitoring to the allocation of the Grid Program and the monitoring to the grid process in every node. The first part should get information that which node the grid process is on earth dispatched to; the second part should monitor the information (CPU usage, Memory usage, Net IO rate-if necessary) of the grid process in every node.After the monitoring to Grid Resource and Program is accomplished, the data from monitoring can be used to do the prediction to the Grid Performance which might have great importance to the allocation of the Resource and the scheduling of the Program in Grid. Currently, there are many methods to predict the resource and the program in a Singleton computer. These methods can also attempt to be applied to the Grid prediction. Artificial Neural Network which has many types is one of the effective methods. BP Neural Network is a typical type of Artificial Neural Network.This issue is part of project"Real-time monitoring and predicting of Grid applications'execution performance"which is funded by National Natural Science Foundation. A suit of methods and prototype are established to: provide real-time execution prediction for Grid resource and applications; direct scheduling system; make the utilization of Grid resources more reasonable; guarantee the threshold value of task's finish time; improve the execution performance of applications.This issue illustrates that how to monitor and predict a Grid which is built with Globus Toolkit. Globus Toolkit is a typical grid midware which is developed by Globus Alliance. It is a software package that used to program grid application. The latest version is 4.0.6. There are some methods andtools provided in Globus Toolkit to implement the function of building, configuration and maintenance of a grid. In this issue, the WS-GRAM which is a task assignment infrastructure in Globus Toolkit is used to program a distributed matrix multiply.Grid Monitoring is implemented by monitoring the distributed nodes, and the result is published to the Grid System using the MDS which is provided in Globus Toolkit to publish information. Meanwhile, these monitoring data can be written into a database, prepared for the prediction program to read. The prediction method is the BP Neural Network. Activation function in the hidden layer is hyperbolic tangent function. After confirm the BP Network parameter by iteration, the error can be controlled in a small range. With the historical data read from the database, BP Network trains. At last, it will construct a model which can be load directly in the prediction.Based on the Grid Monitoring and Prediction implemented in this issue, the research can be intensively proceeded in two aspects. The Grid Monitoring system can be configured as a Web Service. Using the WS-Notification in Globus Toolkit, the subscription can be provided, so that the specificial type of application can subscribe the information which it concerns. The test of other algorithm to predict Grid can be attempted. Finally, a comparison among these algorithms can decide the best one for the Grid Prediction.Grid Technology is a developing Technology. Follow the appearance of the standard of Grid Architecture, there might be many new features in the future, which will lead to the growth of the performance. Compared to mainframe computer, Grid has a low price but high performance. Sometimes, a well-managed Grid system's performance can even exceed mainframe computer. So, the perspective of the Grid is very well. The Grid Monitoring and Prediction module is defined in the OGSA. It is obvious that GGF think much of Grid monitoring and Prediction. Therefore,a more intensive research in this aspect is worthy.