Abstract
This talk provides a high-level overview of two Open MPI collective algorithms implemented using shared memory as a communication medium: barrier and broadcast.  Despite being only a few thousand lines of code, the design and implementation were surprisingly complex, requiring significant amounts of infrastructure extending out into operating system issues such as processor and memory affinity.  This talk will explore the algorithms and related issues such as optimizing for non-uniform memory access (NUMA) architectures, large messages / resource limitations, and the complexity of design and debugging in multi-process shared memory.
 
--
{+} Jeff Squyres
{+} The Open MPI Project
{+} http://www.open-mpi.org/