High-performance cloud computing for exhaustive protein-protein docking

Abstract

Public cloud computing environments, such as Amazon AWS, Microsoft Azure, and the Google Cloud Platform, have achieved remarkable improvements in computational performance in recent years, and are also expected to be able to perform massively parallel computing. As the cloud enables users to use thousands of CPU cores and GPU accelerators casually, and various software types can be used very easily by cloud images, the cloud is beginning to be used in the field of bioinformatics. In this study, we ported the original protein-protein interaction prediction (protein-protein docking) software, MEGADOCK, into Microsoft Azure as an example of an HPC cloud environment. A cloud parallel computing environment with up to 1,600 CPU cores and 960 GPUs was constructed using four CPU instance types and two GPU instance types, and the parallel computing performance was evaluated. Our MEGADOCK on Azure system showed a strong scaling value of 0.93 for the CPU instance when H16 instance with 100 instances were used compared to 50, and a strong scaling value of 0.89 for the GPU instance when NC24 instance with 20 were used compared to 5. Moreover, the results of the usage fee and total computation time supported that using a GPU instance reduced the computation time of MEGADOCK and the cloud usage fee required for the computation. The developed environment deployed on the cloud is highly portable, making it suitable for applications in which an on-demand and large-scale HPC environment is desirable.