[Linux-cluster] Using ping as a heuristic not a great idea?

Martin Waite Martin.Waite at datacash.com
Thu May 20 13:34:07 UTC 2010


Hi,

The arp_ip_target feature of the linux bonded driver appears to perform
similar ping tests, but only when the interface has fallen quiet for a
period.

Could your heuristic work in the same way ?  Only perform the ping if no
traffic has passed through the interface for a while ?

regards,
Martin

-----Original Message-----
From: linux-cluster-bounces at redhat.com
[mailto:linux-cluster-bounces at redhat.com] On Behalf Of Don Hoover
Sent: 20 May 2010 13:48
To: linux-cluster at redhat.com
Subject: [Linux-cluster] Using ping as a heuristic not a great idea?

I know its in a ton of the example cluster configs out there, but we
have had trouble whenever we try to use ping as a quorum heuristic.

What we have seen is whenever a large file transfer happens, the pings
start to get dropped by Linux, and the heuristic test starts to fail,
and the cluster node gets killed.

We have tried tweaking the interval, the count, the ttl, adding a -w
etc.. and nothing really keeps the cluster from killing nodes when they
get really busy network traffic.


I have personally had this heuristic work on small test environments,
but once we put into production on really bust workloads it pretty much
is useless.


It is a good idea in theory to use this because it would help ensure
that in a split cluster situation you would end up with the box which
had network connectivity would win over the one that did not. But...if
it causes your cluster to die periodically its not worth it.


Is this a known issue, but its just never mentioned in any of the
cluster setup examples?

Any one have a similar experience, or have any ideas on how to make it
work in a very busy cluster environment?


Also, this makes me wonder, if I have a two node cluster, with each node
getting 1 vote, the quorum getting 1 vote, and the heuristic getting 1
vote, but set the 'required' to only 2 votes, why would the heuristic
cause a loss of quorum since the node with the quorum disk alone would
have the needed two votes?


--
Linux-cluster mailing list
Linux-cluster at redhat.com
https://www.redhat.com/mailman/listinfo/linux-cluster




More information about the Linux-cluster mailing list