The large gap between sustained and peak performance on current high
performance computing platforms remains a serious problem.  While
there are many causes, two are commonly mentioned: the difficulty in
achieving good single node performance and the complexity of managing
distributed memory code.  Various solutions have been offered to this
problem, including the development of new, high-productivity parallel
languages.  However, to effectively address both of these issues, it
is necessary to address the issue of data locality, for both local
memory hierarchies and distributed memory parallelism.  Experience
shows that this is a very challenging problem for compilers, even in
the most studied cases, which suggests that a more cooperative
approach between the system software and the programmer is required.