A multi-GPU protein database search model with hybrid alignment manner on distributed GPU clusters

In computational biology, technologies of protein database search become more and more indispensable for researchers. As the database size rapidly grows, the search complexity is improving, leading to an excessively long runtime. In this paper, we provide a novel parallel protein database search model based on distributed GPU clusters. The proposed model targets clusters composed of personal computers equipped with graphic processing unit (GPU) cards and can employ computing power of GPUs and CPUs in the cluster nodes. The workload distribution strategy is designed to take the performance of each node into account, which enables the model to work well with inhomogeneous cluster nodes with different hardware configurations. A hybrid alignment approach, an extension to our previous work, is used in our model to make global and local alignments done concurrently. The parallel program is realized with compute unified device architecture (CUDA) parallel framework and Microsoft message passing interface (MS-MPI). In the experiment, the model is tested on a small cluster composed of three personal computers. The results show that the proposed model can achieve a speedup up to 159.89 times over the serial counterpart when searching the Swiss-Prot database.