| Summary: | 碩士 === 國立交通大學 === 資訊工程學系 === 84 === The purpose of this thesis is to study an automatic partition
technique for nested loops, i.e., the computation and data
decomposition method for nested loops. In the NUMA(Non-Uniform
Memory Access time) system, the remote memory access time is
longer than local one, the computation and data decomposition of
the program will strongly impact the system performance, and
they are 謻rrelated. Thus the partition, or decomposition, of
a program's computation and data onto each processor of parallel
machine, incasing the parallelism as well as decreasing the data
communication time, in order to optimize the system performance,
is a key issue of parallel compiler. This thesis is based on Q.
Ning[Ning95]'s method. When the decomposition is not
communication-free, we propose a relax algorithm to improve the
[Ning95]'s method. We first compute the dependence weight, and
relax the data locality constraints according the dependence
weight in the decreasing order, to find out the better pipeline
execution and pipeline comnication decomposition. We also
describe how to estimate the execution time, to evaluate the
decomposition result. This thesis's method has already
implemented on the parallel compiler SUIF[Stan94], and evaluate
the results on a parallel simulation environment.
|
|---|
