(University of Oregon, 1990-01-19) Lo, Virginia M.; Rajopadhye, Sanjay; Gupta, Samik; Keldsen, David; Mohamed, Moataz A.; Telle, Jan
The mapping problem in message-passing parallel processors involves
the assignment of tasks in a parallel computation to processors and the
routing of inter-task messages along the links of the interconnection network.
We have developed a unified set of software tools called OREGAMI
for automatic and guided mapping of parallel computations to parallel architectures
in order to achieve portability and maximal performance from
parallel systems. Our tools include a description language which enables
the programmer of parallel algorithms to specify information about the
static and dynamic communication behavior of the computation to be
mapped. This information is used by the mapping algorithms to assign
tasks to processors and to route communication in the network topology.
Two key features of our system are (a) the ability to take advantage of
the regularity present in both the computation structure and the interconnection
network and (b) the desire to balance the user's knowledge
and intuition with the computational power of efficient combinatorial algorithms.
