|
What Parallel PEST does
Parallel PEST distributes model runs across networked PC's such as are now found in most office environments. Where model run times are large and
adjustable parameters are many, the saving in overall PEST optimization time through the use of Parallel PEST is enormous. Thus Parallel PEST can be used in the calibration of large and complex models where
application of nonlinear parameter estimation techniques would have previously been considered impossible.
Implementation of the highly efficient Gauss-Marquardt-Levenberg parameter estimation algorithm used by PEST requires that derivatives of model
outcomes be calculated with respect to adjustable parameters. Hence for every optimization iteration, at least as many model runs must be undertaken as there are adjustable parameters (sometimes twice as many in the
later stages of the optimization process). By carrying out these model runs simultaneously, each optimization iteration can be completed in a fraction of the time that would have been required had all model runs
been carried out in succession on a single machine.
Use of Parallel PEST is very similar to that of PEST, the major difference being that one extra PEST input file must be prepared informing Parallel
PEST of the networked PC's to which it has access. The model itself can either be installed on each of these PC's or on a single PC to which all of the others have access. Whenever Parallel PEST needs to run the
model, it signals a slave residing on one of the networked machines to carry out that run. Model run management is carried out in such a way that any particular model run is undertaken by the fastest available
machine and all available machines are kept as busy as possible in undertaking simultaneous model runs.
Note that Parallel PEST can also be used to reduce optimization time on a single, dual-processor machine using the same methodology.
|
