Installation guide

GFS runs on each node in a cluster. As with all file systems, it is basically a kernel module
that runs on top of the Virtual File System (VFS) layer of the kernel. It controls how and where
the data is stored on a block device or logical volume. In order for cluster members to
cooperatively share the data on a SAN, GFS is required to coordinate a distributed locking
protocol.
4.6 DLM
The Distributed Lock Manager (DLM) is a cluster locking protocol in the form of a kernel
module. Its ensure that clustered nodes that share storage do not corrupt shared data. DLM
provides a high performance locking mechanism required by the Cluster Logical Volume
Manager (CLVM).
4.7 Fencing
Fencing is essentially the act of isolating a node in a cluster when it is deemed malfunctioning
or otherwise unresponsive, effectively ensuring that it can no longer perform I/O to shared
storage. Fencing typically occurs automatically in order to protect processes on another active
node from modifying the resources during node failures.
Fencing is required because it is impossible to distinguish a true failure from a temporary
hang. If the malfunctioning node is truly down, then it cannot do any damage and so
theoretically no action would be required (it could simply be brought back into the cluster with
the usual join process). However, because there exists the possibility that a malfunctioning
node could believe that the remaining cluster members are themselves malfunctioning, a race
condition could ensue with the possibility of data corruption. Instead, the system must assume
the worst and fence in case of problems.
4.7.1 Power Fencing Systems
The power fencing subsystem allows operational cluster nodes to control the power of failed
nodes to ensure that they do not access storage in an uncoordinated manner. Most power
control systems are network based. They are available from system vendors as add-in cards
or integrated into the motherboard. External power fencing devices are also available. These
are typically rack or cabinet mounted units into which servers are plugged.
4.7.2 SAN Based Fencing
While it is preferable to employ a power fencing solution for the robustness a system reboot
provides, SAN switch fencing is also possible. As with Power Fencing, the need is to protect
shared data. SAN switch fencing works by preventing access to storage at the SAN switch.
4.7.3 SCSI Fencing
SCSI persistent reservations can be used for I/O fencing. All nodes in the cluster must
register with the SCSI device to be able to access the storage. If a node has to be fenced, the
registration is revoked by the other cluster members.
Reference the fence_scsi(8) manpage for further details
16 | www.redhat.com