Abstract
Likelihood-based estimation of nonlinearly approximated DSGE models using particle filter methods is computationally very demanding. Parallel particle filters (PPF) are introduced to speedup the likelihood evaluation. The algorithms, which are based on previous work in the field of signal processing, are presented, implemented and evaluated in the context of distributed memory and the message passing model of programming. The main body of our paper focuses on parallel resampling and on the design of PPF algorithms which minimise inter-processor communication. Attention is restricted to approaches which, by construction, do not affect the statistical properties of the particle filter in any important sense. The algorithms are applied for the estimation of a small scale standard New Keynesian model. Experiments with two PPF algorithms on three different computing clusters illustrate the computational gains that can be achieved using parallel methods. Code profiling reveals that MPI collectives performance is critical in determining the scalability properties of the filters. An overall assessment, taking into account both computational speed and code development effort, suggests that a simple parallel particle filter algorithm is a good choice in the nonlinear DSGE model estimation case.