Distributed Shared Memory Systems and Hybrid MPI/OpenMP Programming
We started Chapter 9 in the book with a closer look at
copper, a cluster of 12 nodes, each node is a shared
memory multicomputer with 32 processors. Closer examination of the
three levels of cache and their particular sizes revealed that it
is worthwhile to make distinguish the 32 processors on one node as
groups of 2 or 8, because of L2 and L3 caches are shared
by 2 and 8 processors respectively.
Discussing the several levels of parallelism brought us to considering
a hybrid MPI/OpenMP approach which has many benefits: fewer
"heavyweight" UNIX processes, and increased speedup and greater
efficiency, as multithreading can keep idle MPI processes busy.
We ended the lecture sketching a hybrid MPI/OpenMP strategy to
implement a parallel Jacobi method.