Closed gfursin closed 5 years ago
Some tricks we to use for working around this in our environment (in case you may find them useful):
Thanks @trws ! Yes, we can try parallelization at some point - we have one search function used across most CK modules so it should be doable. As for using full search - we originally did it because we had to support both Linux and Windows, and didn't have time to specialize it. However, since it's stable now, we can probably provide such a specialization for Linux-based machines.
Hi all! While at SC18, I fixed that with @SteVwonder . I added a CK kernel var to substitute default search directories for soft (/usr, $HOME, etc):
$ ck pull repo:ck-env
$ ck set kernel var.soft_search_dirs=dir1,dir2,...
Just substitute above dirs with something reasonable for your environment. It worked fine and we can even point it to the directory with spack packages!
When preparing a demo workflow for SCC18 (https://github.com/ctuning/ck-scc18) with #CollectiveKnowledge, @spack and @flux-framework during my visit to @LLNL, we've noticed that CK software plugins (http://cknowledge.org/shared-soft-detection-plugins.html) work extremely slowly on NFS, i.e. 3..4 minutes to detect one software dependency.
This is normal since we perform the full and recursive search of all directories by default, however maybe we should provide a flag like --quick to speed up search in such case by looking only in a few most commonly used locations. We can then let a user specify the full path to a given soft if it's not found?
CC @SteVwonder @trws @tgamblin