一种网格并行任务执行时间预测算法

广泛研究了网格环境中并行任务执行时间的预测方法,提出了一种基于案例和人工神经网络的预测算法.该算法充分利用了历史有效信息,尤其是对于同一个任务的多次求解而获得的相似记录,通过建立任务特征模板,将历史任务,即案例进行分类,并利用指数平均值或者线性回归方法进行预测.但是由于网格环境的复杂性,以及有限元求解器在求解问题时的复杂性,导致相似性很难定义,在无法根据模板找到相似性案例的时候,利用人工神经网络预测方法进行预测.该算法在面向多学科应用的模拟与可视化环境中进行了实验,证明该方法具有较好的预测性能.

Run-time prediction algorithm for parallel jobs in grid

Extensive studies on predicting the run-time of parallel jobs in the grid environment were conducted,and a Case and Back Propagation(BP) neural network based Prediction(CBPP) algorithm was proposed.The CBPP algorithm made full use of valid history information,especially similar run-time records for the same job.By constructing the template of job characteristics,history jobs or cases were classified,and then the run time was predicted by exponential average or linear regression method.Because of the complexity of the grid environment and the parallel jobs,it was difficult to define the similarity.The neural network was used to predict the run-time when there were no similar cases in the template library.Experimental results in the Multidisciplinary ApplicationS-oriented SImulation and Visualization Environment(MASSIVE) showed that the CBPP algorithm had satisfactory prediction performance.