We have compared the actual running time of a simple 7-point stencil computation with a 3-D decomposition written in MPI with our simulation of it using BigNetSim. In the program, every chunk of data communicates with its six neighbors in three dimensions. The Jacobi relaxation computation is performed, and the maximum error is calculated via MPI_Allreduce.
The result is shown in Table 2 for a problem with fixed size in all runs. The first row shows the running time of the MPI program on 32 to 256 processors; the second row shows the predicted running time using BigNetSim offline on a Linux cluster. The network parameters are based on Quadrics network specifications. It shows that the simulated execution time is close to the actual execution time.