...

you tell the network upfront, when this request completes, do this next one (very similar to Erez's async executive stuff, but in hardware)

FWIW, Erez's async executive engine doesn't have any request chaining capability. It is actually a different design take on what we had with sock util (mpid/common/sock/iocp) on windows. -Jayesh ----- Original Message ----- From: "Pavan Balaji" <[email protected]> To: [email protected] Sent: Thursday, November 26, 2009 12:00:43 PM GMT -06:00 US/Canada Central Subject: Re: [mpich2-core] MPI on The Register Looks like they are going the Quadrics way, which probably means that they are going to go bankrupt as well :-). I heard about this from the Mellanox folks. If I understood it correctly, it was a firmware based (not true hardware) solution for barrier, broadcast and integer reduce. Btw, Mellanox also came up with a "collective offload" solution for their network adapters, which is nothing more than work request chaining, i.e., you tell the network upfront, when this request completes, do this next one (very similar to Erez's async executive stuff, but in hardware). Also, came to know from Gilad that Oak Ridge had given them a lot of money for the collective offload chip (ConnectX2) under the condition that OSU can't publish the work before they do :-). Rich had some preliminary numbers which were terrible; but I'm sure some workshop will accept it. -- Pavan On 11/26/2009 09:57 AM, Rob Ross wrote:

Some discussion of this "collective acceleration in IB switches": http://www.theregister.co.uk/2009/11/26/sc09_networking_wrapup/

What's the scoop on that?

Rob

-- Pavan Balaji http://www.mcs.anl.gov/~balaji