[jira] [Comment Edited] (MESOS-3521) Support for hardware topology and resources description

["Jonathan Calmels (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MESOS-3521?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14908706#comment-14908706 ] 

Jonathan Calmels edited comment on MESOS-3521 at 9/25/15 9:39 PM:
------------------------------------------------------------------

Well in conjunction with [https://issues.apache.org/jira/browse/MESOS-3366], it would allow agents to describe their hardware in a well known format. This format would be standardized, hence schedulers could act upon it to take better decisions.

Use case scenario:
The agent provides a hook that uses libhwloc and export CPU information (see the example above). A scheduler could look at the resources_info of each offers and launch a task on the agent providing the best clock rate.

Following the same logic, we could export GPUs as a custom resource set, and ask the scheduler to pick say 2 of them that are pairable for a multi-GPU aware application. 



was (Author: exxo):
Well in conjunction with [https://issues.apache.org/jira/browse/MESOS-3366], it would allow agents to describe their hardware in a well known format. This format would be standardized, hence schedulers could act upon it to take better decisions.

Use case scenario:
The agent provides a hook that uses libhwloc and export CPU information (see the example above). A scheduler could look at the resources_info of each offers and launch a task on the agent providing the best clock rate.



> Support for hardware topology and resources description
> -------------------------------------------------------
>
>                 Key: MESOS-3521
>                 URL: https://issues.apache.org/jira/browse/MESOS-3521
>             Project: Mesos
>          Issue Type: Improvement
>            Reporter: Jonathan Calmels
>
> In heterogeneous clusters, tasks sometimes have strong constraints on the type of hardware they need to execute on. The current solution is to use custom attributes to describe resources on the agents.
> While this solution works, the current attribute format is somehow constraining and requires workarounds on both the agent and the scheduler (e.g. Base64 RFC 6920). In addition, we often need to encode the link between an attribute and a resource which is inherently error prone and implementation defined from one scheduler to another.
> I would like to propose a unified format to expose hardware topology and resources information at the agent level.
> This format would effectively extend the following specification: [http://mesos.apache.org/documentation/attributes-resources/]
> {code:title=specification}
> kv : text ":" ( range | kvSet | scalar | text )
> kvSet : "{" kv ( "," kv )* "}"
> info : ( range | text | "*" ) : kvSet
> infoSet : "{" info ( "," info )* "}"
> resourceName : text | "*"
> resourcesInfoValue : text "(" resourceName ")" ":" infoSet
> resourcesInfo : resourcesInfoValue ( ";" resourcesInfoValue )*
> {code}
> {code:javascript|title=example}
> --resources= gpus:{card0, card1};ports:[0-100];cpus:8
> --resources_info= nvidia(gpus): {		
> 			card0: {
> 				uuid: GPU-34e8d7ba-0e4d-ac00-6852-695d5d404f51,
> 				name: GeForce_GTX_980,
> 				path: /dev/nvidia0,
> 				clocks: {
>       					graphic: 1392,
>       					sm: 1392
> 				}
> 			},
> 			card1: {
> 				uuid: GPU-12e457ba-0f4e-bf01-3452-674a5b212c21,
> 				name: GeForce_GTX_970,
> 				path: /dev/nvidia1
> 				clocks: {
>       					graphic: 1392,
>       					sm: 1392
> 				}
> 			},
> 			*: {
> 				driver: 352.39
> 			}
> 	   	 };
> 		 services(ports): {
> 			[0-79]: {
> 				type: daemon,
> 				user: root
> 			},
> 			[80-100]: {
> 				type: web_services,
> 				user: www-data
> 			}
> 		 };
> 		 procs(cpus): {
> 			*: {
> 				name: Intel_i7_6700K,
> 				frequency: 4,
> 				cache: 8
> 			}
> 		 }
> {code}



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