[
https://issues.apache.org/jira/browse/MESOS-2729?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vinod Kone resolved MESOS-2729.
-------------------------------
Resolution: Won't Fix
Actually looking at the DRF paper and sorter tests, it is vital for the DRF
algorithm to keep track of the total resources, not just total allocated
resources. This is because the "dominant" resource is decided based on the
total resources on the box.
Example:
Host with 100 cpus and 10G mem
Framework 1's allocation: 1 cpu and 1G mem
Framework 2's allocaiton: 2 cpus and 1G mem
According to DRF:
Dominant share of Framework 1 is *0.1 mem*, because mem share (0.1 = 1/10) >
cpu share (0.01 = 1/00)
Dominant share of Framework 2 is also *0.1 mem*
But if we only account for total allocated resources:
Dominant share of Framework 1 is *0.5 mem*, because mem share (0.5 = 1/2) > cpu
share (0.3 = 1/3)
Dominant share of Framework 2 is *0.7 cpu*
> Update DRF sorter to not explicitly keep track of total resources
> -----------------------------------------------------------------
>
> Key: MESOS-2729
> URL: https://issues.apache.org/jira/browse/MESOS-2729
> Project: Mesos
> Issue Type: Improvement
> Reporter: Vinod Kone
> Assignee: Vinod Kone
> Labels: twitter
>
> DRF sorter currently keeps track of allocated resources and total resources.
> This becomes confusing with oversubscribed resources because the total
> allocated resources might be greater than total resources on the slave.
> The plan is to get rid of the total resources tracking in DRF sorter because
> it is not strictly necessary. The share of each client can still be
> calculated by doing the ratio of allocation of a client to the total
> allocations.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
