Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
Description
Set 1) "yarn.resourcemanager.monitor.capacity.preemption.total_preemption_per_round" should >= 1 / #cluster-nodes. (For example, cluster has 20 nodes, total_preemption_per_round should be at least 1 / 20 = 0.05. (By default this is 0.1)
This really should be part of stack-advisor.
Set 2) yarn.resourcemanager.monitor.capacity.preemption.natural_termination_factor = 1 (by default is 0.2).
The two configs will affect preemption for large containers like LLAP. We will put these suggestions to documentation, but it will be also good if we can update them in Ambari.