Feature #65

MetaLB: Separate cost/benefit analysis for running LB strategy and performing migration

Added by Phil Miller over 6 years ago. Updated 12 months ago.

Load Balancing
Target version:
Start date:
Due date:
% Done:



MetaLB potentially has two separate decisions to make as execution proceeds:
1. Is current load imbalance bad enough to justify the cost of running the load balancing strategy to compute new assignments?
2. Having obtained new better-balanced assignments, is the improvement they would offer enough to justify the cost of migration they would entail?

We may need to carry object size information (as reported by the sizing PUPper) in the load balancing database to allow estimation of migration costs as soon as decisions are made. Alternately, we can transmit the decisions and compute the expected time cost in a distributed fashion (max over network model, some more detailed contention-based estimate, simulation, etc.).


#1 Updated by Ramprasad Venkataraman over 6 years ago

  • Category set to 6

#2 Updated by Phil Miller about 6 years ago

  • Target version set to Unscheduled

#3 Updated by Sam White over 2 years ago

  • Assignee changed from Harshitha Menon to Kavitha Chandrasekar

#4 Updated by Sam White over 2 years ago

  • Target version changed from Unscheduled to 6.9.0
  • Category set to Load Balancing

#5 Updated by Kavitha Chandrasekar over 1 year ago

Estimation of migration costs is already included in the model based method to select Load balancing strategy. So this issue is somewaht subsumed by

#6 Updated by Sam White about 1 year ago

  • Status changed from New to In Progress

#7 Updated by Juan Galvez 12 months ago

  • Status changed from In Progress to Implemented

#8 Updated by Kavitha Chandrasekar 12 months ago

  • Status changed from Implemented to Merged

The MetaBalancer patch has been merged. The random forest model will be updated to include more training runs. Chare pup size feature will also be included as a hint for migration cost in the updated model.

Also available in: Atom PDF