I85PREP
⌘ K
LPIC-3 High Availability and Storage Clusters study materials
LPIC-3 High Availability and Storage Clusters study resources

LPIC-3 High Availability and Storage Clusters: High Availability Cluster Storage

LPIC-3 High Availability and Storage Clusters · Work through your exam, one topic at a time.

All guide sections

High Availability Cluster Storage

High Availability Cluster Storage

This domain tests shared block storage and cluster filesystems: iSCSI, device-mapper multipath, cluster LVM, DLM, GFS2/OCFS2, and DRBD. The central rule is that only one node may update shared metadata, and no node may mount shared data until locking is active.

Core rules

  • Reach shared LUNs through redundant paths. Enable multipathd, define policies in /etc/multipath.conf, identify WWIDs with multipath -ll, and use /dev/mapper/<wwid> rather than /dev/sdX.
  • Local LVM is unsafe on shared disks. Use cLVM (clvmd, older) or lvmlockd with DLM so metadata changes are serialized. In Pacemaker, order DLM before cLVM before the filesystem.
  • GFS2 and OCFS2 are cluster-aware. GFS2 needs DLM and a lock table (cluster:fs); create one journal per node with mkfs.gfs2 -p lock_dlm -t webcluster:gfs -j 2. Mount through a Filesystem resource, not fstab.
  • DRBD is network RAID1. Pacemaker promotes one node. After split-brain, disconnect both, then use drbdadm -- --discard-my-data connect on exactly one side.
  • Fence first: a node must be fenced before its locks are released.

Worked example

Two nodes share a multipath LUN. Use multipath -ll to get /dev/mapper/mpatha; pvcreate /dev/mapper/mpatha; vgcreate vg_shared; lvcreate -L 100G -n gfslv vg_shared. Create DLM (ocf:pacemaker:controld) and cLVM (ocf:heartbeat:clvm), ordering DLM then cLVM. Format mkfs.gfs2 -p lock_dlm -t webcluster:gfs -j 2 /dev/vg_shared/gfslv; add a Filesystem resource with fstype=gfs2, device=/dev/vg_shared/gfslv, directory=/mnt/gfs. Reasoning: DLM supplies locks, cLVM serializes metadata changes, GFS2 uses the lock table, and two journals cover both nodes.

Traps

Creating a local VG on a shared disk corrupts metadata. Mounting GFS2 manually on both nodes bypasses fencing. Omitting -p lock_dlm creates a local filesystem. Reconnecting DRBD after split-brain without discarding on one side re-creates the split brain. Forgetting to enable multipathd leaves a single path.