Exploiting vector computers for simulation

Computers with vector architecture perform arithmetic much faster than conventional machines but only where the work can be organized into similar operations on all elements of long vectors. Field problems have a suitable structure within the model itself. Most other simulation problems lack such systematic structure but the model as a whole often has to be run many times for either stochastic or optimization reasons. It may then be possible to vectorize across many conceptual replicates of the same program all running at once rather than trying to vectorize within the same program. The conditions allowing this are stated. An advantage of this replication method is that even scalars in the original model become long vectors. A disadvantage is the storage needed, but large amounts are becoming available. An initial test of the method yielded 1000 runs of a stochastic simulation of a control system in the same time as only 20 runs by scalar methods. This demonstrates that simulation runs can now be ‘mass produced’ very cheaply. Possible use of this capability in a very general maximum likelihood method of parameter estimation is discussed.