Repository navigation
Replies: 1 comment
|
That was what I thought but when I recently tested it I noticed that the build will still use the full amount of RAM even if when building a subsampled database. However, the subsampled DBs performed so badly that this argument will be deprecated in favor of estimating the required RAM from the input data directly. There are plans for a low memory version of MEDI tracked in #38. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Due to memory constraints in the cluster I changed params.maxDbSize to 200 in build_kraken.nf (200 GB is as much memory I can allocate for 1 job in the cluster).
If I understand right, this only changes the memory usage for building- or does it affect the size of the kraken database it builds?
All reactions