NPAC Technical Report SCCS-687

Primitives for Problems using Hierarchical Algorithms on Distributed Memory Machines

Sanjay Goil

Submitted December 26 1994


Abstract

In scientific computing a class of problems has been identified that consists of highly structured computations on sets of subdomains that are coupled in a irregular manner. The computational relationship between the subdomain is known only at runtime, which can also change between computation phases. This can lead to load imbalance and unreasonable computation overheads on parallel machines. We present an outline of primitives that are required and the runtime support needed to ease the programming tasks involving dynamic computation structures. We study some applications that exhibit different computational structure but are classified under hierarchical methods and suggest primitives that bind them in a common paradigm. This paper describes the structure of the N-body simulation, Volume Rendering and Radiosity applications. The common primitives they require for an efficient implementation on a distributed memory parallel machine are elaborated.


PostScript version of the paper