Philippe Devloo (UNICAMP - FEC : Faculdade de Engenharia Civil)


OOPAR, an object oriented environment for the development of parallel software

OOPAR consists of a set of classes which makes the implementation of
parallel algorithms easier. Within the OOPAR framework a parallel
algorithm consists of a set of tasks which act on a set of objects
modifying their version. In this sense the OOPAR environment is based
on the dataflow programming concept.

There three basic main interfaces within OOPAR:

Distributed data allows the user to associate a unique id and version
with each submitted data object. The distributed data objects are
managed by a DataManager, which is represented by a single instance on
each processor. The DataManager decides which task objects (explained
below) have access to the data.

Tasks are objects from a class derived from the OOPTask class and
which implement a Execute() interface. The execute method implement
part of a parallel algorithm and transform one or more distributed
data objects. The dependency of the task on the distributed data
objects and their version are declared prior to submitting the task to
the TaskManager. Tasks can be assigned to execute on any processor of
the parallel machine.

The communication manager implements the interface to the
communication library. Tasks and distributed data can be transmitted
to other processors using the serialization mechanism.

OOPAR has been successfully used in the paralelization of a large CFD
code. The main benefits of the use of OOPAR are: gradual
paralelization, a unique code base for both the parallel and serial
code (using #ifdef statements), capability of running the serial code
and parallel code within a single process and a single implementation
for both shared memory and distributed memory architectures.