Optimization of Discrete-parameter Multiprocessor Systems using a Novel Ergodic Interpolation Technique

Modern multi-core systems have a large number of design parameters, most of which are discrete-valued, and this number is likely to keep increasing as chip complexity rises. Further, the accurate evaluation of a potential design choice is computationally expensive because it requires detailed cycleaccurate system simulation. If the discrete parameter space can be embedded into a larger continuous parameter space, then continuous space techniques can, in principle, be applied to the system optimization problem. Such continuous space techniques often scale well with the number of parameters. We propose a novel technique for embedding the discrete parameter space into an extended continuous space so that continuous space techniques can be applied to the embedded problem using cycle accurate simulation for evaluating the objective function. This embedding is implemented using simulation-based ergodic interpolation, which, unlike spatial interpolation, produces the interpolated value within a single simulation run irrespective of the number of parameters. We have implemented this interpolation scheme in a cycle-based system simulator. In a characterization study, we observe that the interpolated performance curves are continuous, piece-wise smooth, and have low statistical error. We use the ergodic interpolation-based approach to solve a large multi-core design optimization problem with 31 design parameters. Our results indicate that continuous space optimization using ergodic interpolation-based embedding can be a viable approach for large multi-core design optimization problems. 1