Multiprocessor-Based Placement by Simulated Annealing

Simulated annealing methods have proven to be particularly successful in physical design applications, but often require burdensome, long run times. This paper studies the design and analysis of standard cell placement by annealing in a multiprocessor environment. Annealing is not static: we observe that the temperature parameter which controls hill-climbing in simulated annealing changes the behavior of an annealing algorithm as it runs, and strongly influences the choice of multiprocessor partitioning strategy. We introduce the idea of adaptive strategies that exhibit different speedups across different temperature ranges. Measured performance of parallel placement algorithms running on a multiprocessor demonstrate practical speedups consistent with our predictions.