[ClusterLabs] Establishing Timeouts
ccaulfie at redhat.com
Mon Oct 10 07:34:04 EDT 2016
On 10/10/16 05:51, Eric Robinson wrote:
> I have about a dozen corosync+pacemaker clusters and I am just now getting around to understanding timeouts.
> Most of my corosync.conf files look something like this:
> version: 2
> token: 5000
> token_retransmits_before_loss_const: 10
> join: 1000
> consensus: 7500
> vsftype: none
> max_messages: 20
> secauth: off
> threads: 0
> clear_node_high_bit: yes
> rrp_mode: active
> If I understand this correctly, this means the node will wait 50 seconds (5000ms x 10) before deciding that a cluster reconfig is necessary (perhaps after a link failure). Is that correct?
No that's not correct. the token timeout is 5 seconds in your example -
because token is 5000mS. the token timeout is always what the value of
token_retransmits_before_loss_const affects the token hold timeout -
which is how long the token is held on a node that has no messages to
send before being forwarded on. So increasing
token_retransmits_before_loss_const changes the number of times per
'token' timeout that the token is actually sent.
In the example above you will see that the token is sent approximately
5000/10 = 500 mS. That's approximate, the value is scaled slightly to
make actual timeouts less likely, and also is affected by messages that
may beed to be sent.
> I'm trying to understand how this works together with my bonded NIC's arp_interval settings. I normally set arp_interval=1000. My question is, how many arp losses are required before the bonding driver decides to failover to the other link? If arp_interval=1000, how many times does the driver send an arp and fail to receive a reply before it decides that the link is dead?
> I think I need to know this so I can set my corosync.conf settings correctly to avoid "false positive" cluster failovers. In other words, if there is a link or switch failure, I want to make sure that the cluster allows plenty of time for link communication to recover before deciding that a node has actually died.
> Eric Robinson
> Users mailing list: Users at clusterlabs.org
> Project Home: http://www.clusterlabs.org
> Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> Bugs: http://bugs.clusterlabs.org
More information about the Users