[jira] [Commented] (MESOS-9780) Improve "picky" framework resource allocation under random sorter.

["xiaowei (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-9780?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16941730#comment-16941730 ] 

xiaowei commented on MESOS-9780:
--------------------------------

which means need a mechanism, to allow mesos framework tell  mesos allocator their resources constraints. For now, mesos offer have two main features: (1) quota management (2) reserve resource on specified agents.  Reserve api is a kind of method, can achieve this target for "picky" framework, but this method will change framework lanuch task behavior, and make too many resources reserved fragments for different roles, and make too many mesos api call for reserve of each applications (e.g.  marathon framework), that will also make allocator resources schedule useless.

> Improve "picky" framework resource allocation under random sorter.
> ------------------------------------------------------------------
>
>                 Key: MESOS-9780
>                 URL: https://issues.apache.org/jira/browse/MESOS-9780
>             Project: Mesos
>          Issue Type: Improvement
>          Components: allocation
>            Reporter: Meng Zhu
>            Priority: Major
>              Labels: resource-management
>
> Picky frameworks are frameworks that are interested in some particular set of resources.
> With the current offer model, such a framework usually keeps declining and filter uninterested offers until accepting an offer that meets its needs.
> While picky frameworks are always prone to performance issues. These frameworks are more likely to experience offer starvation issues under random sorter than the DRF sorter.
> Under DRF sorter, declining offers or Mesos side resource filtering do not affect the framework's dominant resource share. Since other frameworks might get resource allocated at the same time which brings up their shares comparatively, a declined/filtered framework would usually have a higher chance of getting other offers as time goes by (if it keeps declining). This reduces the time such a framework getting what it wants eventually.
> Random sorter, however, is stateless. A decline or filter action has no effect on the chance of a framework getting offers. A framework declining or filtering an offer essentially wastes a shot for nothing. It becomes a truly altruistic act with no perceived gain on the framework side. This makes the random sorter likely to perform poorly compared to DRF in terms of handling picky frameworks.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)