> From: Jeeva Paudel <je...@ualberta.ca>
> Date: August 16, 2010 2:57:56 PM MDT
> To: x10-users@lists.sourceforge.net
> Subject: [X10-users] Place migration and Work stealing policies
> Reply-To: Mailing list for users of the X10 programming language
> <x10-users@lists.sourceforge.net>
>
>
> Hi all,
> I am trying to understand the place migration and work stealing policies
> adopted by X10, and I am still unclear about the following questions:
>
> Does X10 allow work-stealing from remote nodes in a cluster environment, or
> is it confined only to the places resident in the local node?
>
> Under what circumstances, does place migration happen? It occurs to me that
> the runtime system has a choice between place migration and work stealing.
>
> Upon place migration, is the array-data moved to the new node's local memory
> and the array index space remapped?
>
> Thanks,
> Jeeva
>
>
> From: Josh Milthorpe <josh.miltho...@anu.edu.au>
> Date: August 16, 2010 8:06:24 PM MDT
> To: Mailing list for users of the X10 programming language
> <x10-users@lists.sourceforge.net>
> Subject: Re: [X10-users] Place migration and Work stealing policies
> Reply-To: Mailing list for users of the X10 programming language
> <x10-users@lists.sourceforge.net>
>
>
> Hi Jeeva,
>
> work stealing is confined to threads within a single place, and distributed
> array elements do not migrate between places.
> This obviously limits opportunities for dynamic load-balancing when operating
> on distributed arrays. One option to support dynamic load-balancing would be
> to allow data to migrate as you suggest. However this raises some tricky
> issues in addition to the questions you already asked, such as:
> - How much data should be migrated? E.g. if an array element is at the head
> of an object graph, should the remainder of the graph also be migrated?
> - What effect would migration have on activities "in-flight" at the source
> and target places? Operating on an object that is currently being migrated
> may cause problems.
> - What changes are required to the distributed GC algorithm to correctly
> manage migrating objects?
>
> If you are interested in this as an area for research, I'd be delighted to
> discuss it further with you (and the language design team).
>
> Cheers,
>
> Josh
>
Place being an encapsulation of data and activities, I thought the entire
collection of data and activities would be moved to a different computing node
as a result of place migration, isn' t it? Am I missing something here?
I am still awaiting answers to what place migration entails in X10 and when
does the deployment unit decide to migrate a place.
Thanks,
Jeeva
------------------------------------------------------------------------------
This SF.net email is sponsored by
Make an app they can't live without
Enter the BlackBerry Developer Challenge
http://p.sf.net/sfu/RIM-dev2dev
_______________________________________________
X10-users mailing list
X10-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/x10-users
