Charm++ parallel programming system

[Eric Bohm <ebohm AT illinois.edu>]

It should be noted that in the current implementation the delegated
CkMulticastMgr section multicast/reductions will significantly
outperform the 1d array broadcast/reductions at scale.  I really would
not recommend decomposing your problem into column arrays, especially if
you would expect them to span a subset of processors.  The current
(wrong in my opinion) default assumptions of anytime migration support
kills scaling performance for array broadcasts under those conditions.

Filippo Gioachin wrote:
> Alternatively, you can chop your 2-D array into N 1-D array, then each
> can perform its own reduction. If you need to do multicasts that span
> more than one array, you can use a CProxySection spanning all the arrays
> you need.
> Filippo
> 
> On Tue, May 4, 2010 at 08:04, Ramprasad Venkataraman 
> <ramv AT illinois.edu
> <mailto:ramv AT illinois.edu>>
>  wrote:
> 
>     Delegate your sections to CkMulticast if you want section reductions.
> 
>     Ram
> 
> 
>     On May 4, 2010, at 7:36 AM, Phil Miller 
> <mille121 AT illinois.edu
>     
> <mailto:mille121 AT illinois.edu>>
>  wrote:
> 
>     > On Tue, May 4, 2010 at 01:10, Abhishek Gupta 
> <gupta59 AT illinois.edu
>     
> <mailto:gupta59 AT illinois.edu>>
>     > wrote:
>     >> I have a 2-D partitioned matrix and I want to do a reduction along x
>     >> dimension in this 2-D chare Array and this needs to be done for each
>     >> column.  What is the best way to do that ? Using array section
>     >> proxies ?
>     >
>     > Instinctively, I'm inclined to say that there's no baked-in support
>     > for that at present. Depending on how performance critical this
>     > reduction is, you may be best off implementing it in your application
>     > as "everyone sends to (0,y)" and have one element calculate the
>     > reduction, or by manually sending the data up a tree.
>     >
>     > It looks like the libraries in
>     > src/libs/ck-libs/{sparseReducer,sparseContiguousReducer} may serve
>     > your purpose.
>     >
>     > _______________________________________________
>     > charm mailing list
>     > 
> charm AT cs.uiuc.edu
>  
> <mailto:charm AT cs.uiuc.edu>
>     > http://lists.cs.uiuc.edu/mailman/listinfo/charm
>     >
>     _______________________________________________
>     charm mailing list
>     
> charm AT cs.uiuc.edu
>  
> <mailto:charm AT cs.uiuc.edu>
>     http://lists.cs.uiuc.edu/mailman/listinfo/charm
> 
> 
> 
> 
> -- 
> Filippo Gioachin
> PhD Candidate
> Department of Computer Science
> University of Illinois at Urbana-Champaign
> Ph: +1-217-333-4764
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> charm mailing list
> charm AT cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/charm