high performance computing on graphics processing units: hgpu.org

Mining the increasing amount of genomic data requires having very efficient tools. Increasing the efficiency can be obtained with better algorithms, but one could also take advantage of the hardware itself to reduce the application runtimes. Since a few years, issues with heat dissipation prevent the processors from having higher frequencies. One of the answers to maintain Moore’s Law is parallel processing. Grid environments provide tools for effective implementation of coarse grain parallelization. Recently, another kind of hardware has attracted interest: multicore processors. Graphic processing units (GPUs) are a first step towards massively multicore processors. They allow everyone to have some teraflops of cheap computing power in its personal computer. The CUDA library (released in 2007) and the new standard OpenCL (specified in 2008) make programming of such devices very convenient. OpenCL is likely to gain a wide industrial support and to become a standard of choice for parallel programming. In all cases, the best speedups are obtained when combining precise algorithmic studies with a knowledge of the computing architectures. This is especially true with the memory hierarchy: the algorithms have to find a good balance between using large (and slow) global memories and some fast (but small) local memories. In this chapter, we will show how those manycore devices enable more efficient bioinformatics applications. We will first give some insights into architectures and parallelism. Then we will describe recent implementations specifically designed for manycore architectures, including algorithms on sequence alignment and RNA structure prediction. We will conclude with some thoughts about the dissemination of those algorithms and implementations: are they today available on the bookshelf for everyone?