| Summary: | 碩士 === 國立清華大學 === 資訊科學學系 === 83 === Massively parallel and distributed memory machine can provide
enormous computation power. However, the communication
latency has limited the performance. This thesis
describes an implementation of the high-performance
communication library. The library is based on the
theoretical rationale of C. Koelbel "Compiler-Time
Generation of Regular Communication Patterns". This library is
machine independent, and it can run on nCUBE/2, DEC Alpha
farm workstations, and SUN workstations. It uses the
nCUBE message passing library to passing the data in the
nCUBE/2 machine, and it is based on the PVM environment in
the workstations. This library supports the general users,
FORTRAN compiler, and Parallel C++ compiler. It can reduce
the complexity of compiler and enhance the
performance. The current version has been incorporated into
a Parallel C++ environment to support parallel I/O. There are
some other functions in the library. They are the general
functions like gather, scalar, shift, and the special
functions like processor remapping, data moving in the
arrays of the different dimensions, and out of the current
configuration. Our collective communication work involving in
processor remapping and out of current configuration issues
an innovation in this area. It makes the system can work with
an open environment where workstations are dynamically
incorporated into the execution environment.
|
|---|
