Application Level Resource Management in Sequential and Parallel
   			Scientific Codes
                      Andreas Stathopoulos
                 College of William and Mary
                http://www.cs.wm.edu/~andreas

Many research groups nowadays, rely increasingly on medium or even small size 
clusters of workstations to perform their scientific computations.
These clusters usually involve networks with much higher overheads than
traditional MPPs and they are often multiprogrammed. With the emergence
of Grid computing older clusters are often not retired but incorporated
in the computational environment. Relying on a batch scheduler or 
the operating system to manage these resources always provides suboptimal 
solutions.

We explore three different ways that the scientific computing code
can manage these resources itself: multigrain parallelism, 
application-level load balancing, and application-level memory management
to avoid thrashing.  The first two techniques apply on iterative methods
(linear systems and eigenvalue problems), while our application level
memory management scheme is more general. Because the application has 
the ultimate knowledge of its requirements, it can vastly outperform system 
based solutions.