An Approach for Parallel Reading in Multiple Sequence Alignment

We propose an approach for faster file reading of multiple sequence alignment input through the use of MPI-I/O over a subset of MPI cores. The idea is to allow a subset of MPI cores that perform I/O operations and to broadcast locally on individual neighbors, so the code is less sensitive to the stability of the parallel file system. It is achieved by creating a number of subgroups under a global MPI communicator. The size of each subgroup and the buffer size of each reading operation are tuned through the synthetic benchmark. We verify the performance of our approach by comparing it with the traditional way of "sequential file reading and global broadcast", and apply it to the existing MPI version of multiple sequence alignment software ClustalW. In the production run over 8192 BlueGene/Q cores, the current approach provides 6.8 times acceleration of the existing ClustalW-MPI implementation.