qdisk using in redhat 2,3,4 node clusters.

[AVV]

Hi,

I am setting up a 2node, 3 node and 4 node clusters in RHEL 5.3 with cluster suit. I use quorum disk here and can any one explain me well about the below parameters as how to defined with an examples as a 3 node cluster.

1. ============> In this min_score and votes.

2. =============> score

3.cman expected_votes.

I read a lots of links and I could make up the clusters but the basics are yet to cleared. Thanks in advance for your support.

[Steven E. Protter]

Shalom,

3.cman expected_votes.

This can let you weight certain services to run on a particular node. Normally you weight all nodes equally, but there are valid reasons not to do this, such as making services default to more powerful servers.

SEP

[Matti_Kurkela]

SEP:
"cman expected_votes" is a cluster-wide attribute, not a node or service attribute. You're thinking about the "priority" attribute in the failoverdomainnode tag.


AVV:
Please read the qdisk man page (man 5 qdisk) on your systems. It has a detailed explanation of quorumd parameters and heuristics, with example configurations for 2-node and 3-node clusters.

The quorumd heuristics score is calculated locally by each node. It's the sum total of all successful heuristics' scores on that node. If the score is equal or greater than min_score, the node is considered to be "alive". If the node's score is less than min_score, that means the node is faulty and will remove itself from the cluster (= qdiskd will reboot the node).

The qdiskd processes on all "alive" nodes will then communicate with each other using the quorum disk. The node with the lowest node ID (among the "alive" nodes) will be chosen as the master qdiskd. This master node will then grant extra vote(s) for the CMAN cluster quorum calculation. If the node with the master qdiskd dies, a new master is elected.

If one of the nodes does not update its status block on the quorum disk at the configured rate, the master qdiskd can throw it out of the cluster. When this happens, the master qdiskd will both send the node an eviction message through the quorum disk ("Node X, your qdisk updates are late - reboot yourself") and fence it out of the cluster.

CMAN cluster quorum is determined by votes. Each running node will grant (at least) 1 vote for the cluster quorum, and the master qdiskd of the quorum disk system can grant extra votes. The "cman expected_votes" parameter is the number of votes expected when everything is OK.

If the network connections between the cluster nodes fail, the cluster can become partitioned in two or more parts. In these situations, only one of the partitions must continue running: the others must stop because they won't be aware of what the rest of the cluster is doing, and might violate LVM or GFS locks, causing data corruption.

If a cluster becomes partitioned, only the partition that has _more than 50%_ of expected_votes will continue. The nodes in the smaller partition will be "inquorate" (=without quorum) and can only stop all cluster services and wait until the quorate part of the cluster fences them out.

MK

[AVV]

Thanks for the info.