Computing the Phylogenetic Likelihood Function Out-of-Core

The computation of the phylogenetic likelihood function for reconstructing evolutionary trees from molecular sequence data is both memory- and compute-intensive. Based on our experience with the user community of RAxML, memory-shortages (as opposed to CPU time limitations) are currently the prevalent problem regarding resource availability, that is, lack of memory hinders large-scale biological analyses. To this end, we study the performance of an out-of-core execution of the phylogenetic likelihood function by means of a proof-of-concept implementation in RAxML. We find that RAM miss rates are below 10%, even if only 5% of the required data structures are held in RAM. Moreover, we show that our proof-of-concept implementation runs more than 5 times faster than the respective standard implementation when paging is used. The concepts presented here can be applied to all programs that rely on the phylogenetic likelihood function and can contribute significantly to enabling the computation of whole-genome phylogenies.