Аннотация:Parallel program execution on multiprocessor system is affected by many factors. Often it is rather difficult to estimate how program would behave when running on some definite number of processors. Most tools designed for performance and other parallel program characteristics evaluation affect source code of a program. In this paper we consider neural network approach to predict job run time (when using a queue-based submission system). All analysis that is needed for the prediction can be provided without any knowledge about program source code.