You are viewing a plain text version of this content. The canonical link for it is here.
Posted to yarn-issues@hadoop.apache.org by "Carlo Curino (JIRA)" <ji...@apache.org> on 2018/02/15 22:55:00 UTC
[jira] [Commented] (YARN-7834) [GQ] Rebalance queue configuration
for load-balancing and locality affinities
[ https://issues.apache.org/jira/browse/YARN-7834?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16366372#comment-16366372 ]
Carlo Curino commented on YARN-7834:
------------------------------------
The uploaded patch provides an Linear Programming (LP) implementation of this algorithms, leveraging the oljalgo solver (which ships with hadoop already, and contains a pure-java solver, as well as hooks to leverage an external more powerful solver such as Gurobi or CPLEX).
The formulation is designed to:
# Guarantee that all queues will be allocated fully
# Guarantee that none of the sub-clusters is allocated more capacity than it can take
# It maximizes load-balancing (as a primary objective).
# Subject to not impacting load-balancing more than a configurable delta (zero by default), it maximizes queue-to-sub-cluster affinity (as a secondary objective).
The reasons behind 3/4 being in a primary-secondary relationship (instead of a weighted linear combination) is that in our production experience load-balancing is the most concerning issue, secondary of which we aim at optimizing for locality.
> [GQ] Rebalance queue configuration for load-balancing and locality affinities
> -----------------------------------------------------------------------------
>
> Key: YARN-7834
> URL: https://issues.apache.org/jira/browse/YARN-7834
> Project: Hadoop YARN
> Issue Type: Sub-task
> Reporter: Carlo Curino
> Assignee: Carlo Curino
> Priority: Major
> Attachments: YARN-7834.v1.patch
>
>
> This Jira tracks algorithmic work, which will run in the GPG and will rebalance the mapping of queues to sub-clusters. The current design supports both balancing the "load" across sub-clusters (proportionally to their size) and as a second objective to maximize the affinity between queues and the sub-clusters where they historically have most demand.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: yarn-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: yarn-issues-help@hadoop.apache.org