Storage - nthu-ioa/cluster GitHub Wiki

Users have access storage space under the following paths:

/cluster/home (for code and other important small files with light, infrequent I/O);
/data (for many small files or large files, and moderate I/O, including job output and logs).
/lfs/data (for very large files and very high data rates)

These directories are hosted on separate volumes:

/cluster/home is directly attached to our login node;
/data is provided by a separate storage server node, s01.
/lfs is provided by a dedicated Lustre fileserver.

[!TIP]

Space on /lfs is not created automatically, but anyone who needs it is welcome to ask. Please ask either your supervisor or [email protected].

[!WARNING] Never write job output to /cluster/home!

Please avoid long-running/frequent writes and reads from the compute nodes to /cluster/home/!

Heavy IO traffic to home slows down the network and disk access for everyone because it is directly connected to our login node, fomalhaut. The /data disks have a dedicated server that can handle heavier IO loads without disrupting the rest of the network. Please send the output of your compute jobs, including logs and other temporary files, to /data.

[!WARNING]

Never make symbolic links between two different filesystems (e.g. /cluster/home, /data, /data1, /lfs). This can cause all sorts of problems, especially if those files are read or written intensively. If one of the filesystems is /lfs, it is possible to lock up the whole cluster by writing intensively through symbolic links.

Tempted to do ln -s /data/me/my_file /lfs/data/me/my_link? Just don't, please! Whatever you are trying to do, there will be a better way. Please ask.

[!CAUTION] THERE IS NO BACKUP!

Make your own backups (i.e. not on the cluster) of any data you cannot afford to lose.

Although our storage is somewhat robust to the random failure of a few individual drives, it will not survive a catastrophe (e.g. a meltdown of one of our drive arrays, or a serious electrical fault/fire in the machine room).

Scratch space

Some of the compute nodes have fast local scratch drives for temporary storage that could help to speed up IO-intensive jobs. These are not visible on the network, only on the nodes.