[ https://issues.apache.org/jira/browse/YARN-2113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15981780#comment-15981780 ]
Eric Payne commented on YARN-2113: ---------------------------------- [~sunilg], YARN-2113.0009.patch has a problem when {{preemption-order}} is {{USERLIMIT_FIRST}}. Cluster looks like: ||*Queue Capacity*||*Queue Max Capacity*||*Minimum User Limit Percent*||*Default Container Size*||*Vcores Per Container*|| |10GB|20GB|25%|0.5GB|1| With one user taking up the whole queue and cluster, the {{Active Users Info}} section looks like this (which seems accurate except I would have expected the {{Max Resource}} to be 5120, but since it's not active, it probably doesn't matter): ||*User Name*||*Max Resource*||*Used Resource*|| |hadoop1|<memory:20992, vCores:1>|<memory:20480, vCores:40>| When the second user starts, some containers get preempted, but preemption stops early and the {{Active Users Info}} section looks like this (with inaccurate values for {{hadoop1}}: ||*User Name*||*Max Resource*||*Used Resource*|| |hadoop2|<memory:9728, vCores:1>|<memory:2560, vCores:5>| |hadoop1|<memory:5120, vCores:1>|<memory:5120, vCores:10>| - The {{Used Resource}} for {{hadoop2}} is accurate, but {{Used Resource}} for {{hadoop1}} is incorrect. {{hadoop1}} is actually using 17.5GB, while the above says it is using 5GB. Perhaps it's as a result of the following code? {code:title=IntraQueueCandidatesSelector#preemptFromLeastStarvedApp} if (ret && preemptionContext.getIntraQueuePreemptionOrder() .equals(IntraQueuePreemptionOrder.USERLIMIT_FIRST)) { Resources.subtractFrom(userResourceUsage, c.getAllocatedResource()); } {code} > Add cross-user preemption within CapacityScheduler's leaf-queue > --------------------------------------------------------------- > > Key: YARN-2113 > URL: https://issues.apache.org/jira/browse/YARN-2113 > Project: Hadoop YARN > Issue Type: Sub-task > Components: scheduler > Reporter: Vinod Kumar Vavilapalli > Assignee: Sunil G > Attachments: > TestNoIntraQueuePreemptionIfBelowUserLimitAndDifferentPrioritiesWithExtraUsers.txt, > YARN-2113.0001.patch, YARN-2113.0002.patch, YARN-2113.0003.patch, > YARN-2113.0004.patch, YARN-2113.0005.patch, YARN-2113.0006.patch, > YARN-2113.0007.patch, YARN-2113.0008.patch, YARN-2113.0009.patch, > YARN-2113.v0.patch > > > Preemption today only works across queues and moves around resources across > queues per demand and usage. We should also have user-level preemption within > a queue, to balance capacity across users in a predictable manner. -- This message was sent by Atlassian JIRA (v6.3.15#6346) --------------------------------------------------------------------- To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org