Fencing and Stonith

Any computer cluster may be loosely defined as a collection of cooperating computers or nodes. Nodes talk to each other over communication channels, which are typically standard network connections, such as Ethernet.

The main purpose of an HA cluster is to manage user services. Typical examples of user services are an Apache web server or, say, a MySQL database. From the user’s point of view, the services do some specific and hopefully useful work when ordered to do so. To the cluster, however, they are just things which may be started or stopped. This distinction is important, because the nature of the service is irrelevant to the cluster. In the cluster lingo, the user services are known as resources.

Every resource has a state attached, for instance: "resource r1 is started on node1". In an HA cluster, such state implies that "resource r1 is stopped on all nodes but node1", because an HA cluster must make sure that every resource may be started on at most one node.

A collection of resource states and node states is the cluster state.

Every node must report every change that happens to resources. This may happen only for the running resources, because a node should not start resources unless told so by somebody. That somebody is the Cluster Resource Manager (CRM) in our case.

So far so good. But what if, for whatever reason, we cannot establish with certainty a state of some node or resource? This is where fencing comes in. With fencing, even when the cluster doesn’t know what is happening on some node, we can make sure that that node doesn’t run any or certain important resources.

If you wonder how this can happen, there may be many risks involved with computing: reckless people, power outages, natural disasters, rodents, thieves, software bugs, just to name a few. We are sure that at least a few times your computer failed unpredictably.

There are two kinds of fencing: resource level and node level.

Using the resource level fencing the cluster can make sure that a node cannot access one or more resources. One typical example is a SAN, where a fencing operation changes rules on a SAN switch to deny access from a node.

The resource level fencing may be achieved using normal resources on which the resource we want to protect would depend. Such a resource would simply refuse to start on this node and therefore resources which depend on it will be unrunnable on the same node as well.

The node level fencing makes sure that a node does not run any resources at all. This is usually done in a very simple, yet brutal way: the node is simply reset using a power switch. This may ultimately be necessary because the node may not be responsive at all.

The node level fencing is our primary subject below.

Before we get into the configuration details, you need to pick a fencing device for the node level fencing. There are quite a few to choose from. If you want to see the list of stonith devices which are supported just run:

Stonith devices may be classified into five categories:

The choice depends mainly on your budget and the kind of hardware. For instance, if you’re running a cluster on a set of blades, then the power control device in the blade enclosure is the only candidate for fencing. Of course, this device must be capable of managing single blade computers.

The lights-out devices (IBM RSA, HP iLO, Dell DRAC) are becoming increasingly popular and in future they may even become standard equipment of of-the-shelf computers. They are, however, inferior to UPS devices, because they share a power supply with their host (a cluster node). If a node stays without power, the device supposed to control it would be just as useless. Even though this is obvious to us, the cluster manager is not in the know and will try to fence the node in vain. This will continue forever because all other resource operations would wait for the fencing/stonith operation to succeed.

The testing devices are used exclusively for testing purposes. They are usually more gentle on the hardware. Once the cluster goes into production, they must be replaced with real fencing devices.

Stonith is our fencing implementation. It provides the node level fencing.

The stonith and fencing terms are often used interchangeably here as well as in other texts.

The stonith subsystem consists of two components:

The fencing configuration consists of one or more stonith resources.

A stonith resource is a resource of class stonith and it is configured just like any other resource. The list of parameters (attributes) depend on and are specific to a stonith type. Use the stonith(1) program to see the list:

It is easy to guess the class of a fencing device from the set of attribute names.

A short help text is also available:

A dummy stonith resource configuration, which may be used in some testing scenarios is very simple:

An alternative configuration:

This configuration is perfectly alright as far as the cluster software is concerned. The only difference to a real world configuration is that no fencing operation takes place.

A more realistic, but still only for testing, is the following external/ssh configuration:

This one can also reset nodes. As you can see, this configuration is remarkably similar to the first one which features the null stonith device.

The real device configuration is not much different, though some devices may require more attributes. For instance, an IBM RSA lights-out device might be configured like this:

If you haven’t already guessed, configuration of a UPS kind of fencing device is remarkably similar to all we have already shown.

All UPS devices employ the same mechanics for fencing. What is, however, different is how the device itself is accessed. Old UPS devices, those that were considered professional, used to have just a serial port, typically connected at 1200baud using a special serial cable. Many new ones still come equipped with a serial port, but often they also sport a USB interface or an Ethernet interface. The kind of connection we may make use of depends on what the plugin supports. Let’s see a few examples for the APC UPS equipment:

The former plugin supports APC UPS with a network port and telnet protocol. The latter plugin uses the APC SMART protocol over the serial line which is supported by many different APC UPS product lines.

The CRM configuration is left as an exercise to the reader.

Just like any other resource, the stonith class agents also support the monitor operation. Given that we have often seen monitor either not configured or configured in a wrong way, we have decided to devote a section to the matter.

Monitoring stonith resources, which is actually checking status of the corresponding fencing devices, is strongly recommended. So strongly, that we should consider a configuration without it invalid.

On the one hand, though an indispensable part of an HA cluster, a fencing device, being the last line of defense, is used seldom. Very seldom and preferably never. On the other, for whatever reason, the power management equipment is known to be rather fragile on the communication side. Some devices were known to give up if there was too much broadcast traffic on the wire. Some cannot handle more than ten or so connections per minute. Some get confused or depressed if two clients try to connect at the same time. Most cannot handle more than one session at the time. The bottom line: try not to exercise your fencing device too often. It may not like it. Use monitoring regularly, yet sparingly, say once every couple of hours. The probability that within those few hours there will be a need for a fencing operation and that the power switch would fail is usually low.

Apart from plugins which handle real devices, some stonith plugins are a bit out of line and deserve special attention.

http://www.linux-ha.org/wiki/STONITH

http://www.clusterlabs.org/doc/crm_fencing.html

http://www.clusterlabs.org/doc/en-US/Pacemaker/1.0/html/Pacemaker_Explained

http://techthoughts.typepad.com/managing_computers/2007/10/split-brain-quo.html