An improvement of the differential evolution method implementated on GPU

Differential evolution is a very effective numerical optimization method which applied to diverse the set of computationally intensive tasks. Due to the features of the algorithm, it is very suitable for graphics processing unit (GPU). Most of the algorithm stages can be executed independently that corresponds to the basic programming paradigm of GPU (single instruction multiple data). Besides, an algorithm has a regular memory structure for internal data. The use of GPU allows to improve the speed of the algorithm significantly. The analysis of existing implementations of differential evolution method on GPU is performed. Moreover, existing implementations of the differential evolution algorithm on GPU use several unoptimized techniques which restrict effective application of the algorithm to tasks which use multiple optimization procedures. A new computational model that improves current implementations is described. A presented model reduces kernel calls latency due to combining logically-connected kernel into a single global kernel. The algorithm allows to use computational grid which contains single computational block. This approach satisfies requirements of the differential evolution approach to size of population (from five to ten times greater than number of optimized variables) and allows to use GPU internal barrier synchronization techniques. Besides a proposed implementation has regular data allocation in the global memory of GPU that provides coalescence of requests to the slowest global memory thus all threads in warp can read and write information from global memory per single request. Moreover it allows to use Compute Unified Device Architecture (CUDA) streams in very effective manner. In fact, a proposed model can simultaneously execute as much optimization procedure as multiprocessors available on GPU and belimited only by computer capabilities restrictions.