Our user-data is stored in /users on ginseng in a RAID 1 mirror running on two 400 GB SATA disks. Ginseng runs Solaris 10 and uses ZFS as the filesystem for /users. All of our systems NFSv4 mount /users.
We have also explored additional methods for replicating user-data, including AFS, Coda, and DRBD, but have found all to be unusable or problematic.
NFSv3 has been in long standing use by the CSC as well as almost everyone else on the planet. NFSv4 mounts of /users are currently in the works to CSCF. Unfortunately NFS has a number of problems. Clients become desperately unhappy when disconnected from the NFS server. Also previous to NFSv4 there was no way to client side cache, resulting in poor performance with large files.
On November 8, 2007, we experienced a major NFS failure. An analysis of the logs indicated that the fault was likely caused by NFSv4-specific code. As a result, we have returned to mounting with NFSv3.
On March 15, 2008, we transitioned to ZFS.
Each user directory is stored in a separate zfs file system.
To create a user directory:
zfs create users/$USER
To delete a user directory:
zfs destroy users/$USER
To move/rename a user directory:
zfs rename users/$USER_OLD users/$USER_NEW
To disable atime, devices, and setuid:
zfs set atime=off users zfs set devices=off users zfs set setuid=off users
To export over NFS using host-based access-control:
zfs set sharenfs="sec=sys,rw=$ACCESS_LIST,nosuid" users
where ACCESS_LIST may be as a colon-separated list of any of the following:
- hostname (e.g. glucose-fructose.csclub.uwaterloo.ca)
- netgroup (e.g. in LDAP)
- domain name suffix (e.g. .csclub.uwaterloo.ca)
- network (e.g. @22.214.171.124/24)
A minus sign (-) may prefix one of the above to indicate that access is to be denied. 'man share_nfs' has full details.
To make umask work sanely with ACL's:
zfs set aclmode=passthrough users
To make ACL inheritance work sanely:
zfs set aclinherit=passthrough users
Here's what we set this to currently:
The NFSv4 domain is auto-detected by default, although to be safe, you can explicitly set it in /etc/default/nfs:
This documents some important steps that needed to be done once.
You need to create an nfs kerberos principal on caffeine:
sudo kadmin.local addprinc -randkey zfs/ginseng.csclub.uwaterloo.ca ktadd -e des-cbc-crc:normal -k /tmp/ginseng.nfs.keytab
You then need to merge that keytab (using ktutil) into /etc/krb5/krb5.keytab on ginseng.
To query quota:
zfs get quota users/$USER
To set quota:
zfs set quota=$SIZE users/$USER
where $SIZE could be 2.5G, 100M, etc...
To set no quota:
zfs set quota=none users/$USER
It's important to note that quotas include all descendants of a filesystem, including snapshots. This means that if a user deletes a file, but the file is contained in a snapshot, the users available quota will not decrease. To get around this, refquota should be used instead of quota. However, refquota is only supported on OpenSolaris.
Snapshots can be accessed from /users/$USER/.zfs/snapshot/.
To create a snapshot:
zfs create users/$USER@$SNAPSHOT
To delete a snapshot:
zfs destroy users/$USER@$SNAPSHOT
To rename a snapshot:
zfs rename users/$USER@$SNAPSHOT_OLD users/$USER@$SNAPSHOT_NEW
To list snapshots:
zfs list -t snapshot -r users
A rolling snapshot script is available in ~dtbartle/csc/zfs/. It should be configured to run as a root cronjob:
5 2 * * * /usr/local/sbin/snapshot-rotate.py users daily 7 25 2 * * 0 /usr/local/sbin/snapshot-rotate.py users weekly 4 45 2 1 * 0 /usr/local/sbin/snapshot-rotate.py users monthly 4
You should occasional scrub (error-check) the zpool:
zpool scrub users
For small files, ZFS+NFS performance really sucks. You can fix this by enabling zil_disable:
echo 'set zfs:zil_disable=1' >> /etc/system
To see zpool status and statistics:
zpool status -v users zpool iostat -v users