[Users] Re: A question about Node RAM

[Aleksandar Ivanisevic]

Tim Small <tim at seoss.co.uk> writes:

> It also allows you to move nodes around easily and should allow you to
> carry out things like host node kernel updates without bringing down
> containers (using live migration to other HNs) - although we've not
> gotten around to testing this.

I've tested this and its terrible ;) Migration across two drbd volumes
syncing at the same time -- disaster in terms of latency and I/O
speed for the remaining node(s) in the cluster.

> Our machines are in pairs, but really it'd be better to have them in
> something like groups of four, so that when a HN fails, the remaining 3
> HNs each end up running a third of the evicted containers...  This would
> require corosync instead of heartbeat of course (heartbeat supports 2
> nodes only).

Groups of four might work ok provided that the drbd devices are on
separate disks and you are careful always to migrate to an unrelated
machine that doesn't have the standby volume from the source.