Speaker: Dinesh Kaushik, Argonne National Lab

Title:
Understanding the Performance of Hybrid (distributed/shared memory) Programming
Model

Abstract:
With  the availability  of  large scale  SMP  clusters, the  different
software   models   for   parallel   programming   require   a   fresh
assessment.  For  the  physically  distributed  memory  machines,  the
message passing interface (MPI) has been a natural and very successful
software  model. For  another  category of  machines with  distributed
shared memory and non uniform  memory access, both MPI and OpenMP have
been  used with  respectable parallel  scalability.  But  for clusters
with two or more SMPs on  a single node, the mixed software model with
threads within a node (OpenMP  being a special case of threads because
of  the potential  for highly  efficient handling  of the  threads and
memory by the compiler) and MPI between the nodes looks quite natural.
In this talk, we will present our assessment of the hybrid programming
model in the context of a large scale CFD computation involving
irregular memory references.