gearpump-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Karol Brejna (JIRA)" <>
Subject [jira] [Commented] (GEARPUMP-149) Binding tasks to workers
Date Sat, 21 May 2016 12:17:12 GMT


Karol Brejna commented on GEARPUMP-149:

Another rationale: 
I've created a cluster - master and worker on PC (default config), a worker on raspi (less
memory, less slots). 
When deploying a DAG, the master "avoids" creating an executor on raspi. It prefers "stronger"
If we had "binding" mechanism, we could force the task to be ran on raspi.

> Binding tasks to workers
> ------------------------
>                 Key: GEARPUMP-149
>                 URL:
>             Project: Apache Gearpump
>          Issue Type: Improvement
>            Reporter: Karol Brejna
>            Priority: Minor
> Let's assume that we have a cluster with some "specialized" nodes (for example a worker
on IoT device that is responsible for controlling some actuators, dedicated workers to achieve
data locality, some hardware specific - mem/cpu - tasks, etc.).
> There should be a way to "hint" Gearpump to prefer some nodes (workers) over others for
certain tasks (executors). 
> We already have "localities" feature. Developer can add a configuration (when submitting
 the application) to bind a task with a worker:
> {code}
> gearpump {
>   streaming {
>     localities {
>       app1: {            //This is the application name to be submitted
>         workerId: [    //The worker id where to launch Task0_0 and Task0_1
>           TaskId(0,0), TaskId(0,1)
>         ]
>       }
>     }
>   }
> }
> {code}
> The problem is: users have no idea what each Task's task ID is (especially if there are
lots of nodes in a DAG).
> We could introduce more flexible "binding" mechanism, for example labeling tasks and
workers or using some other "awareness" algorithm.

This message was sent by Atlassian JIRA

View raw message