[Linux-cluster] Starter Cluster / GFS

[Gordan Bobic]

Jonathan Barber wrote:
> On 11 November 2010 10:46, Gordan Bobic <gordan at bobich.net> wrote:
>> Jankowski, Chris wrote:
>>> Gordan,
>>>
>>> I did not ask for bonding.  This should work.  I asked for
>>> multiple independent links - different networking interfaces
>>> configured for different IP subnets mapping to different VLANS.
>>>
>>> STP is, these days, run on a per VLAN basis. Having multiple
>>> links in different VLANs protects against important classes of
>>> network failures.  Bonded interface does not do it. This must
>>> be integrated in the clustering software.
>> I don't quite see the point you're making. If your goal is redundant
>> networking, then you can achieve that by having bonded interfaces in each
>> node, and each of the components of the bonded interface should go to a
>> different switch. That will give you both extra bandwidth and a redundant
>> path between all the nodes, which will ensure you don't end up with a
>> partitioned cluster.
> 
> Chris' point is that if the STP has to recalculate (for example if the
> STP root node dies), then having multiple interfaces in the same VLAN
> will not help (if the time taken to recalculate is longer than the
> fencing timeout). But, if he can run the heartbeat across multiple
> VLANs, and the network supports per-VLAN STP, then he lowers the risk
> of both VLANs being affected by the same event and therefore reduces
> the likelihood of a shootout between the cluster nodes.
> 
> Of course, it depends on the topology of the STP domains as to whether
> you are guaranteed to maintain at least one path between nodes in the
> cluster given a STP node failure.

Yes, but your cluster VLAN (the one that's monitored for heartbeating) 
should be isolated, rather than public, so the only nodes on it will be 
the cluster nodes (and probably the SAN). If with that many nodes your 
spanning tree recalculation still takes 40 seconds you have network gear 
that is unfit for purpose anyway.

Gordan