hadoop-yarn-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Eric Payne (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (YARN-5892) Support user-specific minimum user limit percentage in Capacity Scheduler
Date Fri, 07 Jul 2017 12:34:01 GMT

    [ https://issues.apache.org/jira/browse/YARN-5892?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16077246#comment-16077246
] 

Eric Payne edited comment on YARN-5892 at 7/7/17 12:33 PM:
-----------------------------------------------------------

[~sunilg], [~leftnoteasy], [~jlowe]:
Since branch-2 and 2.8 are somewhat different than trunk, it was necessary to make some design
decisions that I would like you to be aware of when reviewing this backport:
- As noted [here|https://issues.apache.org/jira/browse/YARN-2113?focusedCommentId=16023111&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-16023111],
I did not backport YARN-5889 because it depends on locking changes from YARN-3140 and other
locking JIRAs.
- In trunk, a change was made in YARN-5889 that changed the way {{computeUserLimit}} calculates
user limit. In branch-2 and branch-2.8, {{userLimitResource = (all used resources in queue)
/ (num active users in queue)}}. In trunk after YARN-5889, {{userLimitResource = (all used
resources by active users in queue) / (num active users)}}.
-- Since branch-2 and 2.8 use {{all used resources in queue}} instead of {{all used resources
by active users in queue}}, it is not necessary to modify {{LeafQueue}} to update used resource
when users are activated and deactivated like was done in {{UsersManager}} in trunk.
-- However, I did add the activeUsersSet to LeafQueue and all the places it is modified so
it can be used to sum active users times weight.
-- Therefore, it wasn't necessary to create a separate UsersManager class as was done in YARN-5889.
Instead, I added a small amount of code in ActiveUsersManager to keep track of active users
and to indicate when users are either activated or deactivated.
- {{LeafQueue#sumActiveUsersTimesWeights}} should not do anything that synchronizes or locks.
This is to avoid deadlocks because it is called by getHeadRoom (indirectly), which is called
by {{FiCaSchedulerApp}}.
{code}
  float sumActiveUsersTimesWeights() {
    float count = 0.0f;
    for (String userName : activeUsersSet) {
      User user = users.get(userName);
      count += (user != null) ? user.getWeight() : 1.0f;
    }
    return count;
  }
{code}
-- This opens up a race condition for when a user is added or removed from {{activeUsersSet}}
while {{sumActiveUsersTimesWeights}} is iterating over the set.
--- I'm not an expert in Java syncronization. Does this expose {{LeafQueue}} to concurrent
modification exceptions?
--- There is no {{ConcurrentHashSet}} so should I make {{activeUsersSet}} a {{ConcurrentHashMap}}?



was (Author: eepayne):
[~sunilg], [~leftnoteasy], [~jlowe]:
Since branch-2 and 2.8 are somewhat different than trunk, it was necessary to make some design
decisions that I would like you to be aware of when reviewing this backport:
- As noted [here|https://issues.apache.org/jira/browse/YARN-2113?focusedCommentId=16023111&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-16023111],
I did not backport YARN-5889 because it depends on locking changes from YARN-3140 and other
locking JIRAs.
- In trunk, a change was made in YARN-5889 that changed the way {{computeUserLimit}} calculates
user limit. In branch-2 and branch-2.8, {{userLimitResource = (all used resources in queue)
/ (num active users in queue)}}. In trunk after YARN-5889, {{userLimitResource = (all used
resources by active users in queue) / (num active users)}}.
-- Since branch-2 and 2.8 use {{all used resources by active users in queue}} instead of {{all
used resources in queue}}, it is not necessary to modify {{LeafQueue}} to keep track of when
resources are activated and deactivated like was done in {{UsersManager}} in trunk.
-- However, I did add the activeUsersSet to LeafQueue and all the places it is modified so
it can be used to sum active users times weight.
-- Therefore, it wasn't necessary to create a separate UsersManager class as was done in YARN-5889.
Instead, I added a small amount of code in ActiveUsersManager to keep track of active users
and to indicate when users are either activated or deactivated.
- {{LeafQueue#sumActiveUsersTimesWeights}} should not do anything that synchronizes or locks.
This is to avoid deadlocks because it is called by getHeadRoom (indirectly), which is called
by {{FiCaSchedulerApp}}.
{code}
  float sumActiveUsersTimesWeights() {
    float count = 0.0f;
    for (String userName : activeUsersSet) {
      User user = users.get(userName);
      count += (user != null) ? user.getWeight() : 1.0f;
    }
    return count;
  }
{code}
-- This opens up a race condition for when a user is added or removed from {{activeUsersSet}}
while {{sumActiveUsersTimesWeights}} is iterating over the set.
--- I'm not an expert in Java syncronization. Does this expose {{LeafQueue}} to concurrent
modification exceptions?
--- There is no {{ConcurrentHashSet}} so should I make {{activeUsersSet}} a {{ConcurrentHashMap}}?


> Support user-specific minimum user limit percentage in Capacity Scheduler
> -------------------------------------------------------------------------
>
>                 Key: YARN-5892
>                 URL: https://issues.apache.org/jira/browse/YARN-5892
>             Project: Hadoop YARN
>          Issue Type: Improvement
>          Components: capacityscheduler
>            Reporter: Eric Payne
>            Assignee: Eric Payne
>             Fix For: 3.0.0-alpha4
>
>         Attachments: Active users highlighted.jpg, YARN-5892.001.patch, YARN-5892.002.patch,
YARN-5892.003.patch, YARN-5892.004.patch, YARN-5892.005.patch, YARN-5892.006.patch, YARN-5892.007.patch,
YARN-5892.008.patch, YARN-5892.009.patch, YARN-5892.010.patch, YARN-5892.012.patch, YARN-5892.013.patch,
YARN-5892.014.patch, YARN-5892.015.patch, YARN-5892.branch-2.015.patch
>
>
> Currently, in the capacity scheduler, the {{minimum-user-limit-percent}} property is
per queue. A cluster admin should be able to set the minimum user limit percent on a per-user
basis within the queue.
> This functionality is needed so that when intra-queue preemption is enabled (YARN-4945
/ YARN-2113), some users can be deemed as more important than other users, and resources from
VIP users won't be as likely to be preempted.
> For example, if the {{getstuffdone}} queue has a MULP of 25 percent, but user {{jane}}
is a power user of queue {{getstuffdone}} and needs to be guaranteed 75 percent, the properties
for {{getstuffdone}} and {{jane}} would look like this:
> {code}
>   <property>
>     <name>yarn.scheduler.capacity.root.getstuffdone.minimum-user-limit-percent</name>
>     <value>25</value>
>   </property>
>   <property>
>     <name>yarn.scheduler.capacity.root.getstuffdone.jane.minimum-user-limit-percent</name>
>     <value>75</value>
>   </property>
> {code}



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org


Mime
View raw message