Large-Scale Pairwise Sequence Alignments on a Large-Scale GPU Cluster

Partha Pratim Pande from Washington State University presents design of a GPU kernel for performing pairwise sequence alignments for large-scale short sequence datasets generated by next-generation sequencers. He describes a graphics processing unit (GPU) kernel that performs batch NeedlemanWunsch (NW) global alignments where the kernel returns an alignment score divided by the total alignment length for each alignment. The kernel is scalable when used with its MPI-based host software and is capable of achieving high-throughput alignment when run on a CPUGPU cluster. The host software includes a load balancing technique for data sets having sequences of nonuniform lengths. GPUs exploit data-level parallelism by interleaving instructions across a large set of active threads, while CPUs exploit instruction-level parallelism by maximizing the number of in-flight instructions from each thread.