The number of tunable parameters increases dramatically as we push forward to the next node of hyper-NA immersion lithography. It is very important to keep the lithographic process model calibration time under control, and its end result insensitive to either the starting point in the parameter space or the noise in the measurement data. For minimizing the least-squares error of a multivariate non-linear system, the industry standard is the Levenberg-Marquardt algorithm. We describe a distributed computing technique that is natural to the algorithm, and easy to implement in a cluster of computers. Applying this technique to calibrating lithographic process model, we can achieve robust optimization results in nearly constant calibration time.