[
https://issues.apache.org/jira/browse/MESOS-489?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vinod Kone updated MESOS-489:
-----------------------------
Component/s: allocation
> Fix Hierarchical DRF to work correctly
> --------------------------------------
>
> Key: MESOS-489
> URL: https://issues.apache.org/jira/browse/MESOS-489
> Project: Mesos
> Issue Type: Bug
> Components: allocation
> Reporter: Thomas Marshall
>
> The current naive implementation of hierarchical DRF can cause frameworks to
> starve in certain situations.
> For example, imagine a cluster with two resources, cpu and gpu, and two
> roles, r1 and r2. r1 has a single framework f1 which only wants cpu, and r2
> has two frameworks, f2 which only wants gpu and f3 which wants both.
> Assume that each role initially starts with 50% of all resources. f1 will
> return its share of gpus since it doesn't need them, which will then be
> offered to r2 and accepted by f2, putting r2's share above 0.5. Similarly, f2
> will return the cpus it doesn't need which will be offered to and accepted by
> f1. Even if either of those offers is given to f3, it cannot accept them
> since neither contains both resources that f3 needs. This pattern will
> continue until f1 has all of the cpus, f2 has all of the gpus, and f3 is
> starved.
> For more detail, see Ali Ghodsi.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
