Memory usage depends on the database. For a rough estimate of memory requirements for a given db, check the size of the hash.k2d file, which has to fit completely in memory. The manual recommends at least 175GB for the standard database. In my experiene, usage maxes out at the following values:
- Genbank high quality: 64g
- Genbank all: 256g
You should request at least this much memory on a node you plan to do classification on.
Kraken2 is very fast compared to kraken1 or other classification tools. After loading the database in memory the first time (about 5 minutes for genbank hq), subsequet runs proceed very quickly. Using 8 cores on SCG I had the following results with the genbank hq database:
10399748 sequences (2356.32 Mbp) processed in 85.562s (7292.8 Kseq/m, 1652.36 Mbp/m).
That's 85 seconds for a dataset with about 10 million read pairs. Obviously more cores can make things faster for larger datasets.
I've evauluated some of the databases below on sets different metagenomic datasets, and compared the unclassified percentages to our old Kraken1 custom database. See this spreadsheet for a summary of results. Overall, Kraken2 with the high quality database classifies less reads than Kraken1 but much quicker. The full genbank database has increased classification percentages and speed.