Research On Dynamic Task Scheduling Strategy Based On Host Load Prediction

The fast development of cluster of PC makes itself a promising architecture in the computer family. Heterogernetiy,openess and high latency of net, the three features of cluster bring about new challenge in network parallel computing. In which , the task scheduling strategy is a crucial one and has become a hot research spot in the field of parallel processing. Task scheduling problem is how to assign a group of tasks to processors to implement to get the shortest processing time.The design of effective task scheduler will make PCs co-work better so as to fully exploit the computing potential of network of PCs.Because the general forms and several limited forms of task scheduling in the cluster system belong to NP-complete problems, the researchers carried out a lot of its research, proposed a series of dynamic task scheduling algorithm. All of these methods improved dynamic load balancing from some aspect , improved cluster performance.However, the task is dynamically spawned in cluster nodes, leading to the load in nodes will change dynamically, which poses a major challenge to the rational allocation and scheduling of tasks in the cluster environment. If host load is accurately predicted and measured prior to allocation and scheduling of tasks, which will surely improve the task scheduling strategy in cluster system, and increase the efficiency of cluster parallel computing.Predicting host load accurately is the key to achieve efficient load balancing, and it is also an important basis for judging whether there are abnormalities. The traditional method is to treat load prediction as stationary time series, using linear prediction model to predict, the method is simple, but the accuracy is not high. Because host load has non-linear, non-smooth features, and the artificial neural networks do not need to establish the precise mathematical model, have good nonlinear characteristics, which opens up a new avenue for host load prediction. However, neural networks are easy to fall into the local minimum, and have shortcomings of weak overall search capabilities.But the genetic algorithm has good overall optimal search capabilities. Genetic neural networks can combine neural network and genetic algorithm, use genetic algorithms to optimize the the initial weight value of neural network, not only retain the merit of the overall optimal of genetic algorithm, but also have nonlinear characteristics and the rapid convergence of neural networks.Analysing the defects of time series method to predict host load, this paper applies the prediction model based on genetic neural network to predict host load, establishes a prediction model and experimental assessment. For the data-processing problem in area of object-probed, this paper combines host load prediction and dynamic task scheduling effectively, improves the traditional Center Job Dispatcher strategy, and a efficient dynamic scheduling strategy based on PVM platform is put forword, and the realization of model and the related algorithm are given. Finally it uses a use case to test the performance of the algorithm.