GENOPT

The original library does not implement parallel evaluation of individuals. In order to allow for a parallel execution of simulations (as it is done in SIESTA), some parts of the library had to be extended appropriately.

Our optimizer evaluates a whole population in parallel, therefore, the population size determines the scalability (to a cluster of workstations) of the algorithm. The population size should thus be chosen to be no smaller than the total number of CPUs that are used in the optimization task. A too large value for the population size, however, influences the performance of the algorithm in that reproduction occur less often than with smaller population sizes. The population size is chosen at startup of the optimization and is not changed afterwards.

2003-03-27