Next: Acknowledgments Up: Computational Load in Model Previous: Surface type.

Conclusion

The physics computations of the parallel version of CCM2 has been analyzed for load imbalances. We note that both the mean load and imbalance vary with the the type of time step being computed (no-radiation time step, time step with radiation and emissivity and absorptivity calculations, or time step with radiation but without the calculation of emissivity and absorptivity). The diurnal variation of shortwave radiation is the major cause of load imbalance (about 75 percent of the total imbalance during an average time step). This imbalance is due to the additional computation required over the grid points in the day region (receiving solar radiation). Attempts are being made to reduce this imbalance in the parallel model by moving computations from more heavily loaded daylight regions to the less-loaded nighttime processors [4]. Weather patterns (resulting in moist convective instability) are also a major cause of imbalance (about 17 percent of the total imbalance during an average time step). Their occurrence in space and time is not predictable a priori, and although remediation would also involve redistribution of work between processors, the strategy would need to be dynamically adaptive.

The present method of parallelization, which exploits the symmetry about the equator and allocates similar latitudinal ranges of the opposite hemisphere to the same processor, effectively negates the polar day/night asymmetry.

Surface type did not cause major load imbalances in this version of the model, though it was noticeable in the calculations of gravity-wave drag and vertical diffusion. Load imbalances might be more severe, however, if the BATS surface hydrological model or some other coupled model is used.


michalak@
Wed Dec 7 03:37:14 CST 1994