Daniel Templeton created YARN-6676:
--------------------------------------
Summary: ComputeFairShares#computeSharesInternal() is inefficient
and lossy
Key: YARN-6676
URL: https://issues.apache.org/jira/browse/YARN-6676
Project: Hadoop YARN
Issue Type: Improvement
Components: fairscheduler
Affects Versions: 2.8.0
Reporter: Daniel Templeton
Assignee: Daniel Templeton
The algorithm used in {{ComputeFairShares.computeSharesInternal()}} uses a
binary search to arrive at share values that can be calculated more directly.
At 100k jobs, I've clocked the binary search at about 6x slower than a direct
calculation. Also, the binary search doesn't always arrive at the correct
answer. Most notably, it has a tendency to assign more resource shares than
there are resources in the cluster.
I'm cleaning up my implementation, and I'll post a patch when it's ready.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
