Application Of GPU Parallel Techniques In Improved Genetic Algorithm And Molecular Similarity

GPU (Graphics Processing Unit) has numerous cores and outpaces CPU in floating point performance. NVIDIA officially launched CUDA (Computing Unified Device Architecture) as the general purpose computation platform for GPUs in2007. After several years of development, GPUs which support CUDA have markedly improved in performance and functions. GPU parallel computing research has become a hotspot in the field of high performace computing field in recent years.Information entropy-based genetic algorithm introduces information-entropy into evolution process. By constructing the coefficients of narrowing of the searching space for multi-population and then controlling contraction of the solution space, this algorithm can be ensured rapid convergence and the ability of searching optimization solution is enhanced. Based on this, this paper uses CUDA to parallel the genetic algorithm. Mersenne Twister, which is suitable to parallel generating pseudorandom number and all of GA operators including coefficients of narrowing of the searching space are implemented on GPU. We analyze the efficiency and precision of the algorithm by comparing with the serial version. The testing results indicate that this approach has high speedups and maintains reasonable quality.The virtual screening as a kind of computer aided drug design, effectively saves money and shortens the cycle of drug discovery. SHAFTS (SHApe-FeaTure Similarity) is a virtual screening approach based on3D molecular similarity calculation. This approach adopts a hybrid similarity metric which combines molecular shape superposition and chemical feature matching. SHAFTS outperforms several virtual screening methods in hit compounds identification and virtual screening efficiency. This paper uses CUDA and MPI to heterogeneous parallel SHAFTS, We analyze the efficiency and precision of the algorithm by comparing with the serial version. The testing results indicate that the program module on CUDA reaches300speedups and the entire program can reaches10speedups.