Craig Carl wrote:
Gordon -
PGM is still experimental so it will be a while before we would
consider implementing it,
Just because it isn't officially ratified doesn't mean it's not stable.
PGM has been stable enough to be implemented in Windows since around
2002, and there have been libraries for UNIX OS-es for a while, too.
I would also expect to see network issues as a cluster grows.
Performance reducing as the node count increases isn't seen as a bigger
issue?
It would also limit a Gluster cluster to a single
broadcast domain. We would have to find a way to manage these
challenges, do you have any thoughts?
Why not make the transport switchable between unicast and PGM/broadcast?
If you need to cross the broadcast domain boundary, use unicast. This
should probably be selectable per node so that local nodes could be on
broadcast while remote nodes are on unicast (e.g. for an off-site backup
node). This isn't not all that big a deal, though, since you can bridge
networks easily enough - and that would still be more efficient than
amplifying the write bandwidth usage by a factor of the number of nodes.
Gordan
