With the cluster and grid are more popular for scientific research in large scale numerical computing, many middle- wares suiting for those environments have been developed. But two problems restrict their applications. One is that one middleware just can be adapted to one special environment. Which causes developed codes can hardly be reused. Another is some middleware development environments can hardly be used to program for non-computer specialist researchers. To overcome those two drawbacks, this paper presents an encapsulated framework for large scale numerical computing based on YML. This paper also presents a fine-grained parallel Gauss-Jordan algorithm which is used to make some experiments on different middlewares to show the good feathers of the component based framework for large scale numerical computing. Finally comparison between using the framework presented in this paper and using special programmable middleware environment shows that the framework is available for researchers though its overhead comparing to theirs costs used to program.