[jira] [Commented] (MESOS-4429) Add oversubscription benchmark/stress/test framework

["Bartek Plotka (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-4429?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15107556#comment-15107556 ] 

Bartek Plotka commented on MESOS-4429:
--------------------------------------

Let's start a `doc` to define the scope and input/output in details: https://docs.google.com/document/d/1VyjbSXyvxyS95asFjzV5A19B_vcIAiUqgz3y_pMvtPs/edit?usp=sharing (:

Some notes on Serenity framework, [~nnielsen] mentioned:
As you can see it can be controlled via JSON file (quite similar to marathon's REST API input https://mesosphere.github.io/marathon/docs/rest-api.html#post-v2-apps). IMO it gives useful ability to store previous `tasks` and build certain reusable scenarios.

One of the interesting features in this framework is ability to stress slave with different kind of tasks using logic similar to `shares`. For instance you can specify that tasks of type A will be run 3 times more often then tasks of type B  (type A task shares = 3 & type B task shares = 1). As a result the framework will be spawning as many as possible tasks of both types in such defined "distribution". It also support targeting the tasks to particular the host.

It could be a good starting point for us.

> Add oversubscription benchmark/stress/test framework
> ----------------------------------------------------
>
>                 Key: MESOS-4429
>                 URL: https://issues.apache.org/jira/browse/MESOS-4429
>             Project: Mesos
>          Issue Type: Task
>            Reporter: Niklas Quarfot Nielsen
>
> To evaluate the function and quality of oversubscription modules, we could ship a test framework which can:
> 1) Launch on oversubscribed and non-oversubscribed resources in a controlled manner. For example, register as two different frameworks and see that resources from slack resources of one framework can be used by the other.
> 2) Measure time to react for different scenarios. For example, measure the time it takes from slack appearing on a slave to the offer being issued with revocable resources. The time to react for changing usage patterns e.g. time to reclaim oversubscribed resources when regular tasks need them back.
> 3) Count the number of offer rescind, preemptions, etc. to deem the stability of the policy.
> 4) Be able to measure % extra work being able to run.
> 5) Work across different resource dimensions as cpu time, memory, network, caches.
> [~Bartek Plotka] has been working on something similar for Serenity in https://github.com/mesosphere/serenity/tree/master/src/framework which we can reuse as a base.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)