Automatically tuned linear algebra software
SC '98 Proceedings of the 1998 ACM/IEEE conference on Supercomputing
Concurrency and Computation: Practice & Experience
Large-scale maximum likelihood-based phylogenetic analysis on the IBM BlueGene/L
Proceedings of the 2007 ACM/IEEE conference on Supercomputing
Phylogenetic models of rate heterogeneity: a high performance computing perspective
IPDPS'06 Proceedings of the 20th international conference on Parallel and distributed processing
Fine-grain parallelism using multi-core, Cell/BE, and GPU Systems
Parallel Computing
| Hi-index | 0.00 |
The multi-core revolution and the biological data flood that is generated by novel wet-lab techniques pose new technical challenges for large-scale inference of phylogenetic trees from molecular sequence data.We present the first assessment of accuracy and performance tradeoffs between single and double precision arithmetics and the first SSE3 vectorization for computing the Phylogenetic Likelihood Kernel (PLK) which forms part of many state-of-the art tools for phylogeny reconstruction and consumes 90-95% of the overall execution time of these tools. Moreover, the PLK also dominates memory consumption, which means that deploying single precision is desirable to accommodate increasing memory requirements and to devise efficient mappings to GPUs. We find that the accuracy provided by single precision is sufficient for conducting tree searches, but that the increased amount of scaling operations to prevent numerical underflow, even when using SSE3 operations that accelerate the single precision PLK by 60%, generates run-time penalties compared to double precision on medium-sized datasets. However, on large datasets, single precision can yield significant execution time savings of 40% because of increased cache efficiency and also reduces memory footprints by 50%.