In high performance computing large number of applications are executed over heterogeneous devices like CPU, GPU, and combination of more than one type of processors. Application when assigned to particular processor run-time information is recorded which is a measure-describing factor in application for task scheduling. Predicting this information prior to execution is the focus, which further distinguishes implemented approaches. It is not guaranteed that model will provide exact information but the predicted value is not always the similar to actual runtime for a specific task in distributed environment. In proposed work, here is the analysis of predicting runtime in advance so that it is clear and correct which device is suitable for which task in particular application. The predicted time is compared to actual runtime and accuracy is improved by running same task over and over again on the supposed device. The results show that historical information can be stored on the basis of prediction model which gives tolerable accuracy for the further yield of reducing waiting time and increasing processor utilization.