[ 
https://issues.apache.org/jira/browse/YARN-4360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15137865#comment-15137865
 ] 

Carlo Curino commented on YARN-4360:
------------------------------------

[~asuresh], thanks for the comments. I address GreedyReservationAgent 1,2.  
IterativePlanner 1 is not an exact duplication, but I merged the code by 
complicating 
a bit the boolean expression (need to check either direction). I left a comment 
to highlight this. Regarding IterativePlanner (2) I believe that is correct, 
and cleaner 
that what was there before. We are generally using exception to signal a failed 
acceptance. 

StageAllocatorGreedyRLE (1): we need to do it this way to signal via 
side-effects the occupied resources to other runs of StageAllocatorGreedyRLE 
that place other 
stages. E.g., if we have an "AND" type reservation (where 2 stages are allowed 
to overlap), we need to know when we place the second one, that the first one 
already
consumed some resources during a time range t1,t2. This is done through this 
planModifications data structure (passed in by IterativePlanner). 




> Improve GreedyReservationAgent to support "early" allocations, and 
> performance improvements 
> --------------------------------------------------------------------------------------------
>
>                 Key: YARN-4360
>                 URL: https://issues.apache.org/jira/browse/YARN-4360
>             Project: Hadoop YARN
>          Issue Type: Sub-task
>          Components: capacityscheduler, fairscheduler, resourcemanager
>    Affects Versions: 2.8.0
>            Reporter: Carlo Curino
>            Assignee: Carlo Curino
>         Attachments: YARN-4360.2.patch, YARN-4360.3.patch, YARN-4360.5.patch, 
> YARN-4360.patch
>
>
> The GreedyReservationAgent allocates "as late as possible". Per various 
> conversations, it seems useful to have a mirror behavior that allocates as 
> early as possible. Also in the process we leverage improvements from 
> YARN-4358, and implement an RLE-aware StageAllocatorGreedy(RLE), which 
> significantly speeds up allocation.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to