Performance

Performance

Real datasets

Dataset	Threads	Normal run	With --bootstrap 100	Peak memory
22 Listeria genomes, ~3 Mb (Tutorial)	8	3 s	1 min 40 s	220 MB
321 Mycobacterium bovis assemblies, ~4.2 Mb	64	7 s	11 min 44 s	—

Both runs used the default k-mer size (21) and sketch size (10,000), with --nj --me --pcoa.

• Sketching takes most of the time of a normal run. It is parallel (one Mash process per genome) and sketches are reused on the next run into the same output folder, so adding genomes to an existing dataset only sketches the new ones.
• Distances (mash triangle) are multithreaded and take seconds for hundreds of genomes. The number of pairs grows with the square of the number of genomes.
• Bootstrap: each replicate re-sketches all the genomes, so N replicates take about N times the sketching time.

Tree building and PCoA

Time to build each tree from an existing distance matrix (random distances, one process, one run each):

Genomes	UPGMA	Minimum evolution	Neighbour joining	PCoA	Peak memory
1,000	0.1 s	2.3 s	0.1 s	0.1 s	0.2 GB
2,500	0.2 s	1.8 s	1.9 s	0.1 s	0.6 GB
5,000	0.8 s	8.2 s	20 s	0.2 s	1.7 GB
9,000	3.1 s	26 s	1 min 52 s	0.5 s	5.3 GB

• Neighbour joining time grows about with the cube of the number of genomes; above a few thousand genomes, minimum evolution (--me) is faster.
• A distance matrix of n genomes takes n² × 8 bytes of memory: 650 MB for 9,000 genomes. With --bootstrap, each tree-building process (up to 8) holds its own copy.
• For very large datasets, --phylip saves the matrix for dedicated tree programs such as rapidNJ or FastME.

Tips

• Use -t to match the CPUs you have (default: all available CPUs, respecting SLURM and other CPU limits).
• Keep the output folder between runs to reuse sketches; use --clean only when you are done.
• Increase --sketch-size only if you need a finer resolution between very close genomes: time and disk usage grow with it.