[Beowulf] multi-threading vs. MPI

[Christian Bell]

On Sat, 08 Dec 2007, Geoff Jacobs wrote:

> But isn't CAF (and UPC, and Titanium) implicitly message passing for a
> Beowulf anyway? It's attractive because it simplifies the process and
> might be able to optimize communication, but underneath everything it's
> still message passing.

It's message passing to the extent that two processes exchange
"messages" over a network -- but it isn't MPI message passing which
would mean receiver-directed matching and placement of data.  On
clusters with advanced network interfaces, the level of message
passing would translate into low-level RDMA operations whereas an SMP
would implement these "messages" as reads and writes to physically
addressable memory.

What's meant by "message" requires a definition -- one can argue that
invalidating a cache line means sending a "message" to the cache
controller but it's far from what people usually think as MPI-level
messages.  PGAS attempts to provide "better programmability" while
targeting low-level communication primitives that do not involve the
MPI-level message passing baggage (matching, two-sided, pairwise
synchronization).


    . . christian

-- 
christian.bell at qlogic.com
(QLogic Host Solutions Group, formerly Pathscale)