Parallel particle filters for evaluation of the likelihood of DSGE models are implemented and evaluated in a distributed memory message-passing context. In our paper special emphasis is put on the details of the interprocessor communication which is necessary for load balancing in the particle generation step. Parallelisation makes it possible to (i) reduce execution time, (ii) employ more accurate solution methods and/or (iii) improve the accuracy of the filter by increasing the number of particles. Here the main focus is on reducing the time of the filter for an example model with sticky prices and wages. Scalability results are presented for the case when resampling is done at each step of the filter. This may be interpreted as putting an upper bound on execution time and speedup. Still, significant gains in wall-clock time can be achieved, while keeping scalability at reasonable levels. This is of value when the filter is used in Bayesian estimation algorithms. A nice property of the algorithms, when tested on two clusters with switched networks, is that the execution time is largely independent of the particle routing requirement, as measured by the total number of particles that need to be transferred between processors after resampling.