<< Back to shouce.jb51.net

10.2. Fencing Methods

Table 10-1 lists the fencing methods and associated fencing agents that you can use with GFS.

Fending MethodFencing Agent
APC Network Power Switchfence_apc
WTI Network Power Switchfence_wti
Brocade FC Switchfence_brocade
McData FC Switchfence_mcdata
Vixel FC Switchfence_vixel
HP RILOEfence_rib
GNBDfence_gnbd
xCATfence_xcat
Manualfence_manual

Table 10-1. Fencing Methods and Agents

When a GFS cluster is operating, the fencing system executes those fencing agents. Specifically, when using LOCK_GULM, the cluster manager is the master lock_gulmd daemon. The daemon uses the fence_node command to dispatch a fencing agent.

NoteNote
 

Contact an authorized Red Hat support representative if there is a device you wish to use for fencing that is not described in the following sections.

The following sections describe the fencing methods available with GFS.

10.2.1. APC MasterSwitch

APC MasterSwitch power switches are used to power cycle nodes that need to be fenced. The fencing agent, fence_apc, logs into the device and reboots the specific port for the failed node. The fence_apc fencing agent supports nodes with dual power supplies plugged into an APC MasterSwitch. Support for nodes with dual power supplies allows powering down both power supplies in a node, thereby allowing fencing of nodes with dual power supplies. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

Lengthy Telnet connections to the APC should be avoided during the cluster operation. A fencing operation trying to use the APC will be blocked until it can log in.

10.2.2. WTI Network Power Switch

WTI network power switches (NPSs) are used to power cycle nodes that need to be fenced. The fencing agent, fence_wti, logs into the device and reboots the specific port for the offline node. The fence_wti fencing agent does not support nodes with dual power supplies plugged into a WTI NPS. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

Lengthy Telnet connections to the WTI NPS should be avoided during the cluster operation. A fencing operation trying to use the WTI NPS will be blocked until it can log in.

10.2.3. Brocade FC Switch

A node connected to a Brocade FC (Fibre Channel) switch can be fenced by disabling the switch port that the node is connected to. The fencing agent, fence_brocade, logs into the switch and disables the specific port associated with the node.

Nodes with multiple FC paths can have each path disabled. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

Lengthy Telnet connections to the switch should be avoided during the cluster operation. A fencing operation trying to use the switch will be blocked until it can log in.

10.2.4. Vixel FC Switch

A node connected to a Vixel FC (Fibre Channel) switch can be fenced by disabling the switch port that the node is connected to. The fencing agent, fence_vixel, logs into the switch and disables the specific port associated with the node. Nodes with multiple FC paths can have each path disabled. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

Lengthy Telnet connections to the switch should be avoided during the cluster operation. A fencing operation trying to use the switch will be blocked until it can log in.

CautionCaution
 

Red Hat GFS does not support the following Vixel firmware: Vixel 7xxx series firmware versions 4.0 or later, Vixel 9xxx series firmware versions 6.0 or later.

10.2.5. HP RILOE Card

A GFS node that has an HP RILOE (Remote Insight Lights-Out Edition) card can be fenced with the fence_rib fencing agent. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

The fence_rib fencing agent requires the Stunnel software be installed on the system running the fencing agent. Stunnel is required to connect to the HP RILOE card.

10.2.6. GNBD

Nodes that only use GFS with storage devices via GNBD (Global Network Block Device) servers can use the fence_gnbd agent. The agent requires no special hardware. The fence_gnbd fencing agent instructs all GNBD servers to disallow all I/O from a fenced node. When a fenced node is reset and re-imports the GNBD devices, the GNBD servers again allow the node access to the devices. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

NoteNote
 

You must not specify the GNBD fencing agent (fence_gnbd) as a fencing device for GNBD server nodes (nodes that export GNBDs to GFS nodes).

10.2.7. Manual

In the absence of fencing hardware, a manual fencing method can be used for testing or evaluation purposes.

CautionCaution
 

Manual fencing depends on human intervention whenever a node needs recovery. Cluster operation is halted during the intervention; therefore, manual fencing is not recommended for production use.

The manual fencing agent, fence_manual, writes a message into the system log of the node on which the fencing agent is running. The message indicates the cluster node that requires fencing. Upon seeing this message (by monitoring /var/log/messages or equivalent), an administrator must manually reset the node specified in the message. After the node is reset, the administrator must run the command fence_ack_manual to indicate to the system that the failed node has been reset. Recovery of the reset node will then proceed. Refer to Section 6.6 Creating the fence.ccs File and Section 6.7 Creating the nodes.ccs File for information on how to configure with this type of fencing.

10.2.7.1. Usage

fence_ack_manual -s IPAddress

IPAddress

The IP address of the node that was manually reset.