[ https://issues.apache.org/jira/browse/YARN-7600?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
wuchang updated YARN-7600: -------------------------- Description: I know, when AM making container request, if the request container is NODE_LOCAL, then AM will also send out RACK_LOCAL and OFF_SWITCH requests. On the ResourceManager side, if RM successfully assigned a NODE_LOCAL container, I saw that RM canceled RACK_LOCAL and OFF_SWITCH requests because it was a duplicated request and did not need to be allocated any more. However, my problem is that if the **allowRelaxity** for RACK_LOCAL and OFF_SWITCH is true, NODE_LOCAL request is downgraded to RACK_LOCAL , thus NODE_LOCAL's request is allocated with RACK_LOCAL locality, then OFF_SWITCH's duplicated request will be canceled of cource, however I did not see RM cancel the duplicated NODE_LOCAL request, so, won't it lead to NODE_LOCAL request still exist and then be scheduled in the next-round scheduling? was: I know, when AM making container request, if the request container is NODE_LOCAL, then AM will also send out RACK_LOCAL and OFF_SWITCH requests. On the ResourceManager side, if RM successfully assigned a NODE_LOCAL container, I saw that RM canceled RACK_LOCAL and OFF_SWITCH requests because it was a duplicated request and did not need to be allocated any more. However, my problem is that if {code}allow_relaxity = true{code} for RACK_LOCAL and OFF_SWITCH, NODE_LOCAL request is downgraded to RACK_LOCAL and the allocation is successful, ie, NODE_LOCAL's request is allocated with RACK_LOCAL locality, then OFF_SWITCH's duplicated request will be canceled, however I did not see RM cancel the NODE_LOCAL request, so, won't it lead to NODE_LOCAL request still exist and then be scheduled in the next-round scheduling? > Yarn NODE_LOCAL request downgraded to RACK_LOCAL didn't cancel the original > NODE_LOCAL request > ----------------------------------------------------------------------------------------------- > > Key: YARN-7600 > URL: https://issues.apache.org/jira/browse/YARN-7600 > Project: Hadoop YARN > Issue Type: Task > Affects Versions: 2.7.3 > Reporter: wuchang > > I know, when AM making container request, if the request container is > NODE_LOCAL, then AM will also send out RACK_LOCAL and OFF_SWITCH requests. On > the ResourceManager side, if RM successfully assigned a NODE_LOCAL container, > I saw that RM canceled RACK_LOCAL and OFF_SWITCH requests because it was a > duplicated request and did not need to be allocated any more. > However, my problem is that if the **allowRelaxity** for RACK_LOCAL and > OFF_SWITCH is true, NODE_LOCAL request is downgraded to RACK_LOCAL , thus > NODE_LOCAL's request is allocated with RACK_LOCAL locality, then OFF_SWITCH's > duplicated request will be canceled of cource, however I did not see RM > cancel the duplicated NODE_LOCAL request, so, won't it lead to NODE_LOCAL > request still exist and then be scheduled in the next-round scheduling? -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: yarn-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: yarn-issues-h...@hadoop.apache.org