On Fri, Aug 24, 2012 at 04:33:21PM +0300, Constantinos Venetsanopoulos wrote: > Hello team, > > I seem to not have actually understood how hail's algorithm works, > and I would like some insight if possible. > > I have a Ganeti cluster with 13 VM-capable nodes running the latest > Ganeti 2.6.0 stable. > > I see that when adding instances of type 'plain' or 'rbd' all instances > go to the same node, until i hit the IPolicy vcpu or spindle ratio. Then, > another node is picked up and again all instances go to that node, > until this node is full too. Is that the expected functionality of hail?
No, not at all. > Shouldn't it balance the cluster somehow? Yes. I'm not sure what is the problem, especially in case of plain, it should work well; while I can imagine rbd doesn't work well (not enough tested), I'm surprised that plain is broken. > I see that when adding instances of type 'drbd' everything works > as I would expect, meaning that hail picks different nodes of the > cluster for each instance, resulting in a balanced cluster. > > Migration/Failover works well with all templates. Hail picks different > nodes everytime. Hmm, very surprising. I'll test if I can reproduce this with plain. Alternatively, if you can give me: - "hscan -L" output after a few instances have been mis-allocated - gnt-debug allocator - wait, there's a bug in it, so instead the command line you use to add another instance it would help to understand the issue. Thanks for the report! iustin
