[Beowulf] Re: multi-threading vs. MPI

[Kirill Lapshin]

Toon Knapen wrote:
> Many on this list suggest that using multiple MPI-processes on one and 
> the same node is superior to MT approaches IIUC. However I have the 
> impression that almost the whole industry is looking into MT to benefit 
> from multi-core without even considering message-passing. Why is that so?

My understanding is that in cluster setup you will have MP anyway, so 
you have a choice of MP+MT vs MP, i.e. intranode is MP anyway, and on a 
every node you can try to do MT (so called hybrid) or just go with MP 
all the way. The latter is simpler and real world examples of MP+MT 
outperforming MP are remain to be seen.

However if you scope is limited to a single node then MT may be very 
well better choice than MP due to a) universal availability of MT 
libraries (pthread, winthreads) and b) the fact that you can easily and 
efficiently implement MP over MT but not the other way round.

As for multinode scenario I think I am buying the argument that MP wins 
over MT+MP in majority if not all cases.