Performance of virtually all parallel and distributed applications deteriorates when a CPU or a communication link has to be shared, but the extent of the slowdown is application dependent. In our experiments with the NAS benchmarks, the slowdown due to congestion on a single link varied from negligible to 120 percent. Estimation of performance of an application under given network conditions is of central importance for resource selection and resource management in shared computing environments. This paper develops a framework to model the performance of applications with CPU and link sharing. The methodology is based on monitoring the application behavior and resource usage on a controlled testbed. The procedure does not require access to the source code or the libraries. We demonstrate that the performance of applications in simple scenarios of network and CPU sharing can be predicted fairly accurately. For the NAS benchmark suite, we observed that the average error in predicting the execution time in different resource sharing scenarios was in the range of 2-6\% and the maximum error was below 12\%.