[Pacemaker] clusters on virtualised platforms

[Lars Marowsky-Bree]

On 2014-07-17T03:48:51, Alex Samad - Yieldbroker <Alex.Samad at yieldbroker.com> wrote:

> I wonder if there Best practise or how to, on how to run clusters on say VMWare.

We've got many customers running SLE HA (pacemaker/corosync) cluster
inside virtual machines. That works fine.

There are a few obvious caveats. Make sure the VMs are actually running
on different nodes being the most obvious one.

Fencing is another. Typically these environments have shared storage, or
can easily get it via iSCSI (and even easily get 3 devices), so we
recommend the use of "sbd" for fencing.

That - sort of - also implies a network-based quorum that is richer than
merely being able to ping a node.

There are some other concerns that are harder to address. We've seen VMs
"freeze" when the hypervisor deems to take a snapshot or during live
migration. You don't want that to affect the cluster; so set the
corosync token timeout to an appropriate value.


In general, if you can, it makes more sense to run HA closer to the
hardware and not inside the VM - instead, have the HA hypervisor layer
protect the VM as a clustered service. That has many advantages from an
architectural and reliability perspective, not the least of which is
that then HA becomes available for *all* VMs if needed, and the folks
managing their virtualized service don't have to worry about HA
themselves.

Unfortunately, a few customers have choosen hypervisors whose idea of
"HA" and "IO isolation" makes me weep, so they're stuck with running HA
inside their VMs. I consider this a blatant failure of the HVM.


Regards,
    Lars

-- 
Architect Storage/HA
SUSE LINUX Products GmbH, GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer, HRB 21284 (AG Nürnberg)
"Experience is the name everyone gives to their mistakes." -- Oscar Wilde