Table 3.

Time and memory usage of different versions of align-families.py, using different multiple sequence alignment algorithms

version	aligner	time/ memory	CPUs
			1	2	4	8	16	32
0.4	MAFFT	time (seconds)	28,638	15,769	8912	5173	3038	1747
2.15	MAFFT		28,754	14,282	7079	3463	1686	854
	Kalign2		4731	1777	945	600	381	246
0.4	MAFFT	memory (MB)	23,704	12,299	6622	3755	2284	1602
2.15	MAFFT		23,927	12,599	6850	3985	2541	1810
	Kalign2		24,648	23,220	12,408	6668	3781	2327

At low levels of parallelization, Kalign2 made the process over 8 times faster, with a memory usage less than twice as much as MAFFT. The new algorithm sped up the tool between 1 and 2.05x. Naturally, at higher levels of parallelization, the reduction of the job queue bottleneck made more of a difference. Memory usage appeared to not be affected, which is expected due to the small size of the job queue compared with the rest of memory usage. To attempt to disentangle the effects of the job queueing algorithm from all the other changes between 0.4 and 2.15, the two versions were compared with all parameters set as similarly as possible. In both cases, the number of --processes was set to 32 and MAFFT was used as the aligner. Crucially, the --queue-size for the 2.15 version was set to be 32, the same as the number of --processes. This approximates the bottleneck in the pre-2.0 version of Du Novo’s job queueing algorithm. Comparing the median of 3 trials of each, the wallclock time of 2.15 was 27% higher than that of 0.4. This could be because of the higher overhead in the more complicated parallelization algorithm, or other changes between 0.4 and 2.15