Test case: AGARD swept wing

The AGARD test case computes three-dimensional flow around a swept wing at low Mach number - see Yates [#!yates!#]. The mesh used here contains 22K vertices. The MecaGRID is designed for large problems containing thousands or millions of vertices thus this mesh is relatively small for Grid calculations. However, it will give an idea of how small meshes perform on the MecaGRID and permit a comparison between the performance of the implicit and explicit solver options in the AERO-F code. The implicit runs required 128 time steps to obtain a residual convergence of $10^{-6}$. Using the explicit scheme would require several thousands of time steps and a very large CPU time to achieve the same level of convergence. The explicit solver uses a 4-stage Runge-Kutta time scheme. This stringent convergence is not necessary to establish Grid performance. For the Grid performance study, roughly equivalent times for both the explicit and implicit methods were used (on the order of 150 seconds, 256 explicit solver time steps).

Shown in Figure 2 is the non-Globus speedup versus the number of processors on the INRIA-nina and INRIA-pf clusters at INRIA Sophia Antipolis. Even though the mesh is small, a reasonable speedup is obtained with 8-processors (a factor of approximately 7). Mach contours at zero incidence for the AGARD mesh are shown in Figure 3. The view gives the false impression that the wing is at an angle of attack.

Figure 2: MPI speedup on the INRIA clusters using the implicit solver

Figure 3: AGARD wing mesh: Mach contours

In the next sections, the MecaGRID performance is examined for 8 processors and combinations of 8 processors involving the different MecaGRID clusters.