Fig. 11

A Benchmark comparison of the Infrared-based RNARedPrint v2 to the original C++ implementation “v1”. Time is measured as user time; space, as maximum resident set size (RSS). We run the tools on the RNAfold benchmark set [16]. We let both tools generate 1000 samples at fixed weights; note that time and space are strongly dominated by the precomputation phase. To directly compare the implementations of the core algorithms, we run both tools on identical tree decompositions, although Infrared ’s default tree decomposer improves for several instances (including the most expensive one). One observes that the RNARedPrint v2 improves in space and time over the original implementation. Only for very short runtimes, at low treewidth, the C++ implementation has a slight edge, presumably due to less overhead. Both implementation show almost no noticable space increase at low tree widths; however the space requirements of the original implementation increase dramatically for treewidths larger than 10. Due to its extreme space requirement, we didn’t solve the single instance of treewidth 16 with RNARedPrint v1; in other cases, it failed due to a bug. For those instances, we indicate only the performance of version 2 (red crosses). B We used Infrared to compute the treewidths for a set of various phylogenetic networks that were collected from recent studies [52]. Using the Infrared network parsimony model, we count the number of reticulation nodes in the networks and calculate their treewidth. It can be seen that the treewidth rather correlates with the number of reticulation nodes than with network size (number of nodes and edges). Our study on ’real-world’ phylogenetic networks suggests that treewidths are often low in practice; consequently Infrared can effectively compute network parsimony by solving the presented models