A numerical study of the accuracy and simulated parallel efficiency of domain decomposition applied to a finite volume fluid flow solver is presented. Two test cases were studied, developing flow in a channel and flow over a surface mounted rib, each with several subdomain configurations. Three methods to improve accuracy and convergence rates were investigated. An over relaxation scheme was applied at the subdomain interfaces, a subdomain continuity modification was implemented, and the subdomain calculation sequence was modified. All three methods improved the convergence rate. The subdomain continuity modification was necessary for high aspect ratio subdomain configurations to converge. It was found that solutions calculated with low aspect ratio subdomains produced dimensionless velocity differences from identical single domain solutions on the order of 10⁻⁴. Parallel efficiencies ranged from 46% to 154%, with efficiency dropping as more domains are added.

The domain decomposition method was applied to simulate forced convection cooling of an electronic device. The device was modeled as a rib mounted in a channel. Reynold's number, channel blockage ratio and Prandtl number are varied. Thermal conductivities are varied over a range covering most of the commonly use encapsulents. Conjugate heat transfer is used to model the thermal field produced from the heat generating rib. The flowfield produced at a blockage ratio of 0.50 and a Reynold's number of 150 produced two secondary recirculation zones, one on top of the rib and the other on the upper wall after the rib. The other flowfields only produced recirculation zones in front and behind the rib. The maximum temperature, which is the parameter of interest in design applications, was found to be weakly linked to Reynold 's number blockage ratio and

Prandtl number, but for dimensionless thermal conductivity less than 50 the maximum rib temperature is strongly dependent on thermal conductivity. For dimensionless thermal conductivities in excess of approximately 100, this dependency becomes weak.