[Krypton-users] Krypton /home file system load + some advice

[Kent Engström]

Dear Krypton Users,

> we've received reports about the Krypton file system being
> more sluggish than normal.

We found a job that is likely to have been the immediate cause this time
(started at the "correct" time, using a run directory under /home,
etc).

We discussed it with the user, he stopped the job, and the /home load
seems to have dropped to normal levels at once.


We will use this opportunity to give you some advice on file system use
on Krypton:

1) Avoid flat directories with tens of thousands of files in them. That
will be bad regardless of filesystem. Use directory hierarchies.
[Probably not involved in today's problems, but we have been hurt by
that before.]

2) If your job needs to do a lot of manipulation on small files, and it
is contained within a node, and not too large, please use the node-local
scratch file system. It is autocreated when a job starts with the name
present in the $SNIC_TMP environment variable on the nodes. It is
removed when the job exits. The amount of space available is ~ 420 GiB
on normal nodes and ~ 840 GiB on fat nodes.

3) For other situations, use your /nobackup filesystems for your data
processing, including work/run directories.

4) Use /home for configuration files and precious data that you need to
have backed up on tape. That is, copy stuff to /home to protect it.
Or use symlinks to put part of you stuff on /home and part of it on
/nobackup, as appropriate.

5) If you can choose between processing data as a few large files or
many small files, choose the former. Handling huge amounts of small
files will hurt on both /nobackup and /home, to varying degrees.

-- 
Kent Engström, National Supercomputer Centre
kent at nsc.liu.se, +46 13 28 4444