Charm++ parallel programming system

[Sameer Kumar <sameerk AT us.ibm.com>]

I think we disabled rzv messages for
some reason, Eric should know best.  So compile charm with -DOPT_RZV=1
 to get full six link throughput.  Regular charm++ messages ~1MB
in size should achieve close to full throughput. 

-Sameer.

Edgar Solomonik <solomon AT eecs.berkeley.edu>
Sent by: ppl-bounces AT cs.uiuc.edu

12/30/2011 04:08 AM

To	PPL <ppl AT cs.uiuc.edu>, charm AT cs.illinois.edu
cc
Subject	[ppl] exploiting multi-link bandwidth of on Blue Gene

Hello,

I've been working on an algorithm that tries to exploit the bandwidth of
every link on a torus network.  Basically, each node sends data to
neighbors in each dimension of the torus, rather than a single dimension. 
The target is to achieve injection bandwidth rather than link bandwidth
on torus networks (e.g. on BG/P injection bandwidth is 6x link bandwidth,
and on BG/Q injection bandwidth is 10x link bandwidth).

I have been able to employ this idea to get a significant performance imporvement
on BG/P for an implementation of matrix multiplication that uses MPI_Put. 
I also wanted virtualization so I implemented the same algorithm in Charm++. 
The topology-aware mapping works and the Charm++ version performs almost
as well as the original MPI version.  However, I've been unable to
get the Charm version to saturate multiple links on BG/P.  I even
tried running with virtual node mode on BG/P and having chares on different
processes within the same node send messages along different torus directions. 

To summarise, I want to have a chare send multiple simultaneous messages
to chares located on torus neighbors in different directions.  Can
I use CkDirect or some other technique to achieve the above goal in Charm++? 
Basically, I need an asynchronous (one-sided) send implementation 
on BG/P.

Thanks,
Edgar_______________________________________________
ppl mailing list
ppl AT cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/ppl