Some performance results for a connection machine implementation of the boundary contour system

Implementation techniques and performance results for an implementation of the BCS (boundary contour system) on the Connection Machine model CM-2 are presented. In particular, the authors are able to duplicate their completion and end-cut experiments on a 128×128 network with four layers, 12 orientations, and a 19×19 F/G-kernel size, with between one and two orders of magnitude speedup over a serial implementation on an IBM 3090. The authors achieve average elapsed times of 65 s per full BCS step, of which approximately 11 s represents work done in the CM-2 back end. The elapsed time is dependent on the time-sharing load of the front-end computer and may vary by as much as 50%. This speedup derives from efficient implementations of the F/G-kernel convolution and the underlying ordinary differential equation solver. To attain even higher performance, it is necessary to understand the CM-2 back end at the microcode level