[ClusterLabs] Two node cluster without fencing and no split brain?
ccaulfie at redhat.com
Wed Jul 21 05:27:02 EDT 2021
On 21/07/2021 09:50, Frank D. Engel, Jr. wrote:
> OpenVMS can do this sort of thing without a requirement for fencing (you
> still need a third disk as a quorum device in a 2-node cluster), but
> Linux (at least in its current form) cannot. From what I can tell the
> fencing requirements in the Linux solution are mainly due to limitations
> of how deeply the clustering solution is integrated into the kernel.
> There is an overview here:
An interesting document (if rather out of date now in some areas at
least). I used to work on VMS up to late V5 (I now work on corosync, but
also started the original linux DLM) and always wanted to get Linux
clustering up to that standard. There are several reasons why that
wasn't really possible.
Firstly Linux has a write-back disk cache which makes sharing disks
between machines MUCH harder, and limits a lot of what you can do. Many
of the limitations of GFS2 seem (to me) to be caused by this. VMS - at
least when i was a sysadmin - wrote straight to devices or via dedicated
intelligent controllers that were also a shared cluster resource
(HDC50/75s in my day). I see VMS6 introduced a "Cluster-wide virtual I/O
cache" which sounds like something Linux could do with. But good luck
getting that merged ;)
Secondy, you are right, we never really got kernel buy-in. The original
Sistina CMAN (which I wrote) was a kernel module, partly because I hoped
that we could get better integration that was (it never happened) and
partly we thought we might need to avoid too many kernel/usermode
context switches for GFS
Thirdly, and mainly, I got the impression that people didn't, in the
main, want that type of cluster. As that document correctly points out,
most Linux clusters are simple(ish) two node failover clusters.
Therefore because also of 1 and 2 above we pursued the path we have.
I would love for Linux to have the cluster capabilities that VMS had in
the 80s/90s but not only is is a massive amount of work, you'd need buy
in from a lot of people who don't really see the point of it.
> I am wondering how much of what OpenVMS does could be integrated into
> Linux in the future to simplify the HA clustering situation. This is one
> thing OpenVMS currently does FAR better than any other platform I've
> come across, so it is likely there is still much to be learned from it.
> On 7/20/21 6:45 PM, Digimer wrote:
>> On 2021-07-20 6:04 p.m., john tillman wrote:
>>> Is it possible to configure a two node cluster (pacemaker 2.0) without
>>> fencing and avoid split brain?
>>> I was hoping there was a way to use a 3rd node's ip address, like from a
>>> network switch, as a tie breaker to provide quorum. A simple successful
>>> ping would do it.
>> Quorum is a different concept and doesn't remove the need for fencing.
>>> I realize that this 'ping' approach is not the bullet proof solution
>>> fencing would provide. However, it may be an improvement over two nodes
>> It would be, at best, a false sense of security.
>>> Is there a configuration like that already? Any other ideas?
>>> Pointers to useful documents/discussions on avoiding split brain with
>>> node clusters would be welcome.
>> (note: currently throwing a cert error related to the let's encrypt
>> issue, should be cleared up soon).
> Manage your subscription:
> ClusterLabs home: https://www.clusterlabs.org/
More information about the Users