Caviness

The Caviness cluster was built in late 2018 according to a rolling upgradeable design strategy. Like previous community clusters, over time additional nodes will be added. But starting in the sixth year of its life, nodes exiting their warranty period will be replaced and the infrastructure surrounding them reused.

The larger amount of reusable infrastructure is afforded by the Open Compute Project (OCP) design of the cluster. Penguin Computing's Tundra design uses centralized DC power conversion in the racks to make node sleds as small and efficient as possible. Remove the old sleds and insert the new.

Many Engineering research groups own capacity in Caviness.

The design of Caviness is similar to previous community clusters.

There are two private ethernet networks in the cluster. A dedicated 1 Gbps network carries management traffic (remote power control of nodes, console access, etc.). A dedicated 10 Gbps network carries all data traffic (NFS, job scheduling, SSH access) to the nodes.

A 100 Gbps Intel Omni-path network also connects all nodes. The OPA network carries Lustre filesystem traffic as well as most MPI internode communications.

Each rack of compute equipment added to Caviness is designed to add storage capacity to the cluster:

• Lustre Object Storage Targets (OSTs) and Servers (OSSs)
• NFS server

The addition of OSTs/OSSs increases the aggregate capacity and bandwidth of the /lustre/scratch filesystem. Individual NFS servers provide distinct capacity and bandwidth but do not aggregate with existing capacity or bandwidth — in short, they're just “more space.”

A discussion of each distinct kind of storage available to users is found below. General usage scenarios for each include:

• Home directory
  • User's personal software builds and installs
  • VALET package definitions for those software installs
• Workgroup directory
  • Software builds/installs used by one or more members of a workgroup
  • VALET package definitions for those software installs
  • Members' data sets, job working directories, job summary directories
  • Project data sets, job working directories, job summary directories
• Lustre scratch
  • Temporary storage of large data sets
  • Temporary storage of job working directories

Each user is granted a home directory with a 20 GiB limit (quota). Typically users will build software in their home directory. The relatively low quota often means that users cannot (and should not) submit computational jobs from their home directories. Home directories are mounted at the path /home/<uid_number>, where <uid_number> is a user's Unix UID number (an integer value, use the id command to determine it).

The home directory is the location of a few important files and directories:

The use of ZFS snapshots as backup copies of a home directory is discussed elsewhere. In general, the editing of .bashrc and .bash_profile is discouraged, especially for the alteration of the PATH and LD_LIBRARY_PATH environment variables.

Each workgroup that purchases capacity in the cluster receives a workgroup directory with a quota in proportion to its level of investment in the cluster: the more compute capacity purchased, the more space granted. Workgroup directories are mounted at the path /work/<workgroup-id> on all nodes in the cluster.

The typical layout of a workgroup directory often includes:

None of these directories are mandatory, but they do tend to make management of a workgroup's resources easier. In particular, the fact that /work/<workgroup-id>/sw/valet will be automatically added to the VALET search path means workgroup users do not need to alter VALET_PATH manually, in their .bashrc/.bash_profile, or in their job scripts.

The use of ZFS snapshots as backup copies of a workgroup directory is discussed elsewhere.

The /lustre/scratch file system is a high-speed parallel file system accessible from all nodes in the cluster. Users/groups are free to create their own top-level directories under /lustre/scratch and are responsible for:

• setting appropriate permissions to restrict access accordingly
• removing files/directories that are no longer needed to keep capacity available over time

The total capacity can be checked using the lfs df command:

 lfs df
UUID                   1K-blocks        Used   Available Use% Mounted on
scratch-MDT0000_UUID  2989410560   977815040  2011593472  33% /lustre/scratch[MDT:0]
scratch-OST0000_UUID 51093819392 12543377408 38549933056  25% /lustre/scratch[OST:0]
scratch-OST0001_UUID 51093630976 11691081728 39402001408  23% /lustre/scratch[OST:1]
scratch-OST0002_UUID 51093489664 12198661120 38894347264  24% /lustre/scratch[OST:2]
scratch-OST0003_UUID 51093485568 12000202752 39092807680  23% /lustre/scratch[OST:3]
 
filesystem_summary:  204374425600 48433323008 155939089408  24% /lustre/scratch

Note that this command displays both aggregate capacity and the capacity of each OST and MDT (MetaData Target) component of the file system. Users can determine their current occupied Lustre scratch capacity:

$ lfs quota -u $(id -u) /lustre/scratch
Disk quotas for usr 1001 (uid 1001):
     Filesystem  kbytes   quota   limit   grace   files   quota   limit   grace
/lustre/scratch  313298       0       0       -      78       0       0       -

Likewise, capacity associated explicitly with a workgroup can be checked:

$ lfs quota -g $(id -g) /lustre/scratch
Disk quotas for grp 1001 (gid 1001):
     Filesystem  kbytes   quota   limit   grace   files   quota   limit   grace
/lustre/scratch  267784       0       0       -      16       0       0       -