Table 6.

CPU Time and Memory Usage for Computing Distances, Trees, and Support Values

Program	Support	COG2814		PF00005		16S rRNA
		h	GB	h	GB	h	GB
FastTree 1.0	None	0.06	0.16	0.52	0.3	16.3	2.4
FastTree 1.0	Local 1,000	0.08	0.16	0.56	0.3	17.3	2.4
Log-corrected distancesa		0.05	0.13	0.71	2.8	33.1	49.9
Maximum likelihood distancesb		138	0.72	≈ 3, 000	—	≈ 5, 000	—
Clearcut 1.0.8c	None	0.06	0.22	1.44	5.2	≈ 28.6	≈ 52
RapidNJ 1.0.0c	None	0.05	2.2	≈ 0.9	≈ 55	≈ 22.1	≈ 549
FastME 1.1c	None	0.51	4.2	≈ 12.5	≈ 105	≈ 138	≈ 1, 000
QuickTree 1.1c	None	0.24	0.16	22.7	2.9	≈ 1, 500	≈ 47
QuickTree 1.1d	Boot 100	63.5	0.71	≈ 104	≈ 15.5	≈ 105	≈ 254
BIONJc	None	32.9	0.44	≈ 820	≈ 10.9	≈ 105	≈ 110
PhyML 3e	Approximate	> 1,000	9.5	—	—	—	—
	likelihood ratio test
RAxML VI 1.0f	None	> 1,000	0.70	—	—	—	—
Consenseg	Boot 100	1.09	0.36	118	9.4	≈ 3, 700	≈ 94

NOTE.—aLRT, approximate likelihood ratio test.

^aThe time to compute the distances between all N² pairs of sequences in the alignment, as implemented by the authors, and the space required to store the N(N – 1)/2 distinct entries of the distance matrix. For nucleotide sequences, these are the same as Jukes–Cantor distances.

^bFor protein sequences, we used PHYLIP's protdist and default options (JTT model, no variation of rates across sites). For nucleotide sequences, we used PHYLIP's dnadist with the F84 model and gamma-distributed rates.

^cThese timings include half of the time to compute N² log-corrected distances because the method requires a distance matrix but each pair of sequences only needs to be considered once.

^dUsing QuickTree's built-in implementation of %different distances and of global bootstrap.

^eFor best performance, we used no variation of rates across sites.

^fFor best performance, we used no variation of rates across sites and the fast hill-climbing option (-f d.). For an initial topology, we used the BIONJ tree.

^gThis does not include the time to compute the resampled trees.