Performance Modeling And Prediction Of Grid Applications

Connecting distributive located and heterogeneous high performing computing resources, storage resources, data resources and other special resources, Grid Computing, realizes dynamic resources optimized combination. It achieves high performing associated computation to solve magnitude application problems. The Virtual Super Computer is composed of all kinds of resources with distinctions among hardware, operation systems, organizations and regions masked by Grid infrastructure, shared and managed transparently and efficiently by different users and organizations. Grid computing has powerful data processing capability and is capable of making full use of idle computing resources through internet.Grid architecture is the core technique of Grid, the skeleton and soul of Grid. A reasonable system structure can be used to design and build Grid, ensure the capability of Grid. Grid should accept all kinds of equipments, any one that follows Grid's rules could join in Grid. Unified and standard interfaces are provided, which are independent from the conditions of accessing equipments. User does not need specific training or knowing of technical details to use Grid. Resources in the Grid can enter or exit without influence to the use of entire system.Resource monitoring is the base of resource prediction. Computing Grid needs real-time, precise, dynamical performance information. Monitoring begins with sensor, any program that generates performance monitoring events with timestamp can be used as a sensor. Sensor collects various dynamical information, such as CPU load, free memory space and network state. Resource prediction model is generated according to monitoring information, then time series of resource prediction is acquired to predict the using of resource in a certain duration.Resource prediction (such as load information of applicable host) could be converted to runtime prediction of application. Referring to the dynamical change of resource, this method can adapt the dynamical and heterogeneous environment of Grid. Time series is provided in real time. Measurement andprediction of resource are separated from applications, and they are periodical. Applications that share resources could share resource prediction.Application performance prediction predicts application performance in the Grid environment using dynamical and heterogeneous performance information. Prediction must combine with task's resource demand in order to figure out predicted execution time as direction of scheduling.Corresponding methods and tools are needed in finding suitable prediction model and using this model to establish a swift predicting system with low overhead; then the methods are used in host-load prediction to predict task's execution time with host-load prediction; finally, prediction information is produced and published to serve Grid users and scheduling programs.This paper is part of project"Real-time monitoring and predicting of Grid application 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 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.Firstly, this paper introduces Grid basic knowledge including Grid infrastructure, also refers to serious of basic predicting methods including serial self-regression model, Gray Predicting Model and Markov Predicting Model. Then it expatiates about Grid Resources predicting methods in detail, such as Regression Predicting Algorithm based on Services Compression and Iterative Algorithm. At the end, Grid Application Performance Predicting Methods are compared, Modeled and Experimented with.To introduce Services Oriented Regression Predicting Algorithm, we refer typical program code first, and then talk about definition of services oriented types and distilling algorithm. After that, a dynamic forecasting method based on BP neural network was introduced to predict the runtime of tasks in a grid. In practice, the runtime of a task is affected by many factors, so linear model cannot describe system features well. The neural networks have well non-linear mapping capability, fast parallel processing ability, powerfully self-learning and self-organizing ability and so on. BP networkshas simple structure, powerful simulation ability and are convenient to implement etc. BP networks have been widely used for evaluation and prediction, expert system, fault diagnosis and so on. With a view to several factors like task scale and host load, we present a dynamic predicting system to predict the runtime of task on hosts using improved BP algorithm. The system achieves a high level of accuracy for exemplar applications. The method is a useful and effective for resource performance prediction and has many advantages, including simple structure, fewer hidden units, faster learning, real-time and on-line predicting, better adaptability, and lower overhead, convenient use and so on.In the future, work from following aspects-conformity of performance predicting information, Embed predicting services and constituting unique rules-should be finalized. Thus, the Grid Application Performance Prediction would be optimized. Future tasks could be expanded from following aspects: 1) Combining all application predicting information such as prediction of memory, bandwidth and hard disk space needed for running task, and building API for showing dynamic predicting report to users. 2) Embedding predicting services into Grid Toolkits, such as Globus Toolkit 4.0. 3) The bottle neck of system performance is unavoidable when Grid nodes increase and Grid scale expands, thus, one unique criterion including code criterion, communication criterion and Grid service criterion should be set up. Then the performance of Grid could be improved firmly.Grid is a new technology, related standard and criterion is being established and improved. It must become new resources fully communicated and shared network environment stepping after Internet and Web.