[ https://issues.apache.org/jira/browse/MYRIAD-249?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15704397#comment-15704397 ]
Tao Jie commented on MYRIAD-249: -------------------------------- https://github.com/apache/incubator-myriad/pull/97 > Should set NodeManager vcores more flexibly > ------------------------------------------- > > Key: MYRIAD-249 > URL: https://issues.apache.org/jira/browse/MYRIAD-249 > Project: Myriad > Issue Type: Bug > Affects Versions: Myriad 0.2.0 > Reporter: Tao Jie > > Today we set resource of NodeManager by configuration like: > {code} > profiles: > zero: # NMs launched with this profile dynamically obtain cpu/mem from > Mesos > cpu: 0 > mem: 0 > small: > cpu: 2 > mem: 1024 > medium: > cpu: 4 > mem: 4096 > large: > cpu: 10 > mem: 12288 > {code} > cpu/mem here is request for Mesos. We launch NodeManager and set > {{nodemanager.resource.cpu-vcores}} and {{nodemanager.resource.memory-mb}} as > cpu/mem once resource is allocated. However the meaning of vcores in YARN is > not extremely the same with cpu in Mesos. In Yarn, we may set vcores to 12 > when physical cpu is 6, and it would be converted to real cpu when request > for vcores. Also in yarn, requested vcores must be integer, as a result each > task would take at least one vcore(actually not necessary one real cpu). > We could have one more field configuration that multiply real cpu in mesos to > vcores in yarn. Perhaps set vcores directly in configuration? > I am freshman of Myriad and Mesos, please correct me if I am wrong. -- This message was sent by Atlassian JIRA (v6.3.4#6332)