[ https://issues.apache.org/jira/browse/FLINK-6309?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Zhuoluo Yang updated FLINK-6309: -------------------------------- Description: Currently in {{PlanFinalizer}}, we travel all the job vertices to calculate the consumer weights of the memory, and then assign the weights for each job vertex. In case of a large job graph, e.g. with multiple joins, group reduces, the value of consumer weights will be very high and the available memory for each job vertex will be very low. I think it makes more sense to calculate the consumer weights of the memory at the job vertex level (after chaining), in order to maximize the usage ratio of the memory. was: Currently in {{PlanFinalizer}}, we travel the whole job vertexes to calculate the memory consumer weights, and then assign the weights for each job vertex. In a case of a large job graph, e.g. with multiple joins, group reduces, the consumer weights will be high and the usable memory for each job vertex will be very low. I think it makes more sense to calculate the memory consumer weights in job vertex level (after chaining) to maximize the memory utility. > Memory consumer weights should be calculated in job vertex level > ---------------------------------------------------------------- > > Key: FLINK-6309 > URL: https://issues.apache.org/jira/browse/FLINK-6309 > Project: Flink > Issue Type: Improvement > Components: Optimizer > Reporter: Kurt Young > Assignee: Xu Pingyong > > Currently in {{PlanFinalizer}}, we travel all the job vertices to calculate > the consumer weights of the memory, and then assign the weights for each job > vertex. In case of a large job graph, e.g. with multiple joins, group > reduces, the value of consumer weights will be very high and the available > memory for each job vertex will be very low. > I think it makes more sense to calculate the consumer weights of the memory > at the job vertex level (after chaining), in order to maximize the usage > ratio of the memory. -- This message was sent by Atlassian JIRA (v6.4.14#64029)