Parallelizing BLAST and SOM Algorithms with MapReduce-MPI Library

Most bioinformatics algorithms are developed in a serial form due to a fast pace of changes in the subject domain and the fact that many bioinformatics tasks can be parallelized as collections of serial jobs communicating at the file system level (High-Throughput Computing, HTC). Recently, a MapReduce-MPI library was made available by Sandia Lab to ease porting of a large class of serial applications to the High Performance Computing (HPC) architectures dominating large federated resources such as NSF Tera Grid. Using this library, we have created two open-source bioinformatics applications. The first one addresses a problem of adapting existing complex and highly optimized serial bioinformatics algorithm to HPC architecture in a minimally invasive way. We built a parallel BLAST implementation that calls the high-level methods of unmodified NCBI C++ Toolkit. We demonstrated scaling for up to 1000 cores on TACC Ranger cluster when processing the sufficiently large input datasets. Using unmodified NCBI Toolkit ensures that the results are compatible across the multitude of settings in the original serial algorithm, and that future versions of the upstream code can be easily integrated. The second application is a Self-Organizing Map (SOM) machine-learning algorithm, popular in bioinformatics applications such as metagenomic binning. The nature of the SOM requires a global synchronization step with a frequency that necessitates the use of an HPC environment. Our implementation of the "batch SOM" uses a mix of MapReduce-MPI and direct MPI calls and scales to 1000 cores as well. This allows easy processing of datasets with a size that is out of range of the serial SOM implementations. Both implementations are available in the open source at http://andreyto.github.com/mgtaxa/.