On Fri, 8 Mar 2013 11:05:01 +0100 Lars Marowsky-Bree <l...@suse.com> wrote:
> On 2013-03-07T21:34:47, Vladimir <m...@foomx.de> wrote: > > > All resources are only able to run if they are distributed in the > > right combination. A working example could like: > > The algorithm is somewhat simplistic, which has the advantage of being > fast. It works "quite well" in scenarios where there's a number of > nodes available and differences between resources are not so large, > but it won't always find the "optimal" solution. > > Optimal placement (rucksack problem) is NP-complete, so all realistic > algorithms implement heuristics. > > In theory, the case you describe is probably one that can be solved > easily enough, but the current one doesn't. We do accept patches ;-) > > (If someone is looking for a master's thesis topic, transforming the > constraints into linear equations and applying an appropriate > optimization function for location scores/utilization and finding a > discrete solution using one of the available libraries is probably the > way to go.) I understand the main problem. I just hoped there is another approach. > > Is there a possibility to configure it in another way? > > You can influence the placement of resources using resource > priorities, but of course that gets a bit hackish for larger > configurations and not exactly automatic. > > If you know the work packaes exactly, you can also use collocation > sets. That's quite feasible for a low node count, which is where the > current algorithm is least effective. Collocations were exactly what I try to avoid. The setup is planned to get >15 resources (and an upper limit is not defined). I think it would get pretty hard to consider all possible collocations, especially if a kind of automated deployment is regarded. Using larger sets of collocation makes the configuration more difficult to read an especially to maintain. > > Furthermore if there is a lack of configured cores and thus a > > resource cannot be started I don't see any log messages or crm_mon > > output. > > This part is normal, yes. You don't get an error message if -inf > colocation constraints prevent a resource from being placed, either. Ok, I see but I'm looking for a possibility to monitor such states to be informed if a resource can't be started because of lack of provided utilization. Does anybody has an idea about that issue? crm_simulate got mentioned by Michael but I'm not sure wether it is the right tool for a monitoring purpose. > > dc-version="1.1.6-9971ebba4494012a93c03b40a2c58ec0eb60f50c" \ > > There have been some fixes/improvements to utilization based placement > since. But I'm not sure if they'd help you. I'm not sure either. Furthermore I "have to" use packages provided by the "regular" ubuntu 12.04 repository. _______________________________________________ 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://bugs.clusterlabs.org
