Conclusions

It was relatively easy to get a 2-D finite difference migration program running on the Connection Machine. A speed significantly greater than that of a well-vectorized Convex algorithm was obtained immediately. Going from this initial success to the ``expected'' level of performance was more difficult and time-consuming. Careful consideration of how a problem maps onto the Connection Machine hardware is required in order to get the best performance from the machine.

Imaging is the slow step in the migration because of the extensive inter-processor communication required. Perhaps more work will discover a way to speed it up.

I hope that the lessons I learned in this exercise will help others to understand better the process of porting a program to the Connection Machine, and give a rough idea of the work needed to bring a particular algorithm ``up to speed''. Hopefully some of these lessons apply directly to finite-difference migration problems that others might attempt to port to the CM.