Implementation of a parallel genetic algorithm for floorplan optimization on IBM SP2

A Multi-Selection-Multi-Evolution (MSME) scheme for parallelizing a genetic algorithm for floorplan optimization is presented and its implementation with MPI and its experimental results are discussed. Our experimental results on a 16 node IBM SP2 scaleable parallel computer have shown that the scheme is effective in improving performance of floorplanning over that of a sequential implementation. The parallel version could obtain better results with more than 90% of probability. Given 1000 second wall clock time, our parallel program could reduce both chip area and maximum path delay by more than 8% with 8 processors and 12% with 12 processors. Parallel computing can also speed up the evolution process so that there could be higher probability of obtaining a better solution within a given time interval.