[mpi3-coll] July/August telecon

[Torsten Hoefler]

Hi Adam,
> The two key areas I'd like to work with our users on here at LLNL are 
> non-blocking and neighbor-exchange (topological) collectives.  I could 
> perhaps also look into the benefits of persistent collectives.  Our 
> users would like to start using non-blocking collectives asap -- they 
> are excited about this.  I suspect we'll be able to show some real value 
> for the neighbor-exchange collectives on our clusters, as well.  And 
> given IB's need to pin memory, we may also be show performance gains 
> with persistent collectives.
Great. I'd like to support you as much as I can.

> For non-blocking collectives, should I start with libNBC?  Where can I 
> get the latest copy of it?  I know you have done some work with 
> neighbor-exchange collectives, too.  Do you have a similar library for that?
yes, the newest version of LibNBC is on the webpage. It offers IB
optimizations:
http://www.unixer.de/research/nbcoll/libnbc/

This version also implements shift and neighbor exchange (on a graph). I
am working on experimental versions of neighbor reductions right now.
However, it is really hard to achieve overlap in real applications and
there are many many things that one can adjust. So before your users get
disappointed, do the theoretical analysis of how much can be overlapped,
run profiles and run the NBCBench benchmark to see if your network
supports overlap well. I can help with all of this.

Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Indiana University    | http://www.indiana.edu
Open Systems Lab      | http://osl.iu.edu/
150 S. Woodlawn Ave.  | Bloomington, IN, 474045-7104 | USA
Lindley Hall Room 135 | +01 (812) 855-3608