[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Changes to the find available node algorithm

Hi Everyone,

   Want to let everyone know we are making changes to the find available algorithm.  The new algorithm will be the previous secondary algorithm.  It is as follows.

- Get all the districts
- Get all the the nodes that will respond quickly
- And among those nodes pick the ones with plenty of available capacity that are in districts with available uids.
- Then randomly pick one of the max(4, 1/4 of the total valid nodes) best fit nodes.

The bad of this change:
  - The best case for finding an available node is a little slower.

The good:
  - Anyone managing OpenShift has less to think about when adding capacity.  Today you have to add enough capacity in new districts so those districts don't get overwhelmed by new gear creates.  With this change that's no longer the case.  You can add a new node or a new district without being concerned about the new district being overwhelmed.  You can also avoid ever doing district compaction by never creating more nodes in districts than you target running in the long term.  If you are targeting 5% active and can run up to 100 gears per node.  That would mean you want 3 nodes per district.  If you just never add more than 3 nodes per district you will never have to compact districts.  You can do the calculation of how many nodes per district you want as:  6000/(max_active_gears in resource_limits.conf for the node_profile of the district (Ex: 100) / target active percentage (Ex: .05 (5%))).  The number 6000 will likely rise at some point but that's the current restriction based on available ports.
  Note: it is still the case when adding a node that you want to add the node to a district with available uids.
  - The distribution of new gears across nodes within your system will always be ideal so that not too much load is being added to particular nodes that happen to be the most available on just district or node capacity
  - The worst case for finding an available node is now faster than before.

Let us know what you think or if you have any concerns about the change.



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]