On 08/24/11 17:38, Vladislav Bogdanov wrote: > 23.08.2011 12:19, Gao,Yan wrote: > [snip] >> When allocating every resource, we compare the capacity of the nodes. >> The node has more remaining capacity is preferred. This would be quite >> clear if we only define one kind of "capacity". While if we define >> multiple kinds of "capacity", for example: >> >> If nodeA has more cpus remaining, while nodeB has more ram and storage >> remaining -- nodeB has more capacity. >> > > Thanks a lot for clarifying this. > So, each remaining capacity parameter values are compared between nodes, > and each node gets "1" pseudo-score if it has more capacity, and "0" > otherwise. Then that pseudo-scores are summed and that sums are > compared. Nodes with greater sum win. > > Did I understand you correctly? Basically, yes.
> >> If nodeA has more cpus and storage1 remaining, nodeB has more ram and >> storage2 remaining -- They have the equal capacity. Then the first >> listed node is preferred. > > Wouldn't it be better from the load-balancing point of view to use > round-robin placement instead of just picking first listed node? Ah, forgot that. We do load-balancing if they have the equal capacity. You can see that from sort_node_weight(). > >> >>> >>> Also could you please describe algorithm used for placement in little >>> bit more details, so I would ask less stupid questions. Does it have >>> something related to linear programming (f.e. >>> http://en.wikipedia.org/wiki/Simplex_method)? >> An optimal solution definitely requires some mathematical optimization >> method. Though we don't use one so far. It's somewhat complicated to be >> introduced and combined with the current allocation factors. > > Thank you for information. Do you have any plans implementing this? Not yet > I > did some linear optimization works when studied at university long ago, > and it wasn't too complicated to actually code the algorithm. That would be nice! > Although > it was very hard to understand what exactly all that methods do and how > to apply them to a real-life task. Indeed. At least We'll need to think of how to handle that if there are collocation constraints, even though we can override the location/stickiness thing to seek for "optimal". Regards, Gaoyan -- Gao,Yan <y...@suse.com> Software Engineer China Server Team, SUSE. _______________________________________________ Pacemaker mailing list: Pacemaker@oss.clusterlabs.org http://oss.clusterlabs.org/mailman/listinfo/pacemaker Project Home: http://www.clusterlabs.org Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
