Exploration and Exploitation of Data Parallelism for Distributed Computing Environment
Date Issued
1998-07-31
Date
1998-07-31
Author(s)
DOI
872213E002043
Abstract
Writing an effcient program for massively
parallel machines (MPP) requires domain
knowledge about message-passing and data
distribution concepts. In order to efficiently
utilize the computing resources, efforts are
devoted to eliminating the communication
overhead. We propose applying a data
replication approach together with data
alignment to solve the problem. In the
proposed method, data arrays can be replicated
in the whole or in part into all processors or
subset of processors. The existing alignment
approach is not well suited to this kind of
problem. Therefore, an abstract computation
array is proposed as the template, and data
distribution is achieved through partitioning
the template among processors. A data array
can be fully replicated to eliminate the spatial
data dependences if there exists no true
dependence across iterations on the data array.
Arrays with loop-carried data dependences can
be partly replicated if we properly dispatch the
iterations to processors. Performance
evaluation is carried out in a Cray T3D and the
results show that the proposed approach is
effective under a restricted loop model.
Subjects
Massively parallel processors
Data Alignment
Distributed memory
Program
partitioning
partitioning
Data partitioning
Data alignment
Publisher
臺北市：國立臺灣大學電機工程學系暨研究所
Type
report
File(s)![Thumbnail Image]()
Loading...
Name
872213E002043.pdf
Size
43.38 KB
Format
Adobe PDF
Checksum
(MD5):c2fe4a8b9ecf102afb4a67d860820ed8