Research On Interactive Parallelization For CFD

As a hot research topic in parallel computing area, automatic parallelization is attracting more and more researchers. Much progress in exploiting coarse-grain parallelism has been made in recent years, but application results are still disappointing, with many programs achieving little or no speedup while executed in parallel. Our team has accumulated much experience after several years of intensive research on the automatic parallelization for CFD(Computational Fluid Dynamics). The automatic parallelizer NPUPAR we developed for YH-III super computer has shown promising result. Considering our achievements, National Laboratory of Parallel and Distributed Processing(PDL) at National University of Defense Technology recommended that we carry out the research on interactive parallelization to get better preference. This paper is just supported by the Defense Science Foundation to carry out the research on the interactive parallelization for CFD.The following is the achievements of the paper:1. The paper proposes a Domain-Computing Model for the parallelization of CFD programs. This model fits structural features of CFD programs better comparing with existing ones, including Frame-Iteration Model and Field-Loop Model that we proposed earlier. The Model enables the in-depth analysis, which helps to ensure the correctness and to improve the efficiency of the resultant parallel programs.2. The paper introduces domain dependency for domain operations. Domain operation is a kind of uniform operation over a data set, which contains data parallelism. This character of domain operation makes domain dependency suitable to the parallelization of CFD programs. Meanwhile, taking advantage of the flexibility of domain operations, parallelizer can carry out not only global analysis on frame-iteration scale and field-loop scale, but also in-depth and detail analysis on various smaller scales. This helps the parallelizer to generate more efficient parallel programs.3. [Wolfe96] proposed a method that recognize induction variables using FUD chains. This paper presents a new method, called EFUD method, to recognizeHIreduction and reduction variables. This method reconstructs FUD chains using the concepts of domain and domain dependency. The new graphs after reconstruction is called EFUD chain. EFUD method is just tailored to the character of CFD program, whose dependencies are generally uniform. Even array reduction, as well as complex scalar reduction, can be recognized by this method.4. The paper studies the way to find communications that should appear in **parallel programs based on domain dependency. Conditions of communication and formulae for computing the region of those data that require communication are given.Then the paper present a strategy to optimize communication using communication tokens. Communication token is a tuple including communication domain, communication direction, and so on. Here, communication domain is the term to represent above data area. Communications can be optimized manually in interactive mode more easily. The strategy reduces the amount and times of communication by merging the tokens, and minimizes the number of synchronizations by placing communications beside the statement which possesses most tokens.5. In order to make interaction more productive, the paper constructs a tree framework to organize data used by parallelizer and corresponding analyzing programs. We call this kind of framework as object tree. Object tree simplifies the procedures of data location, integration, and conversion. Moreover, the tree framework facilitates incremental analysis in order to shorten parallelizer 's responding time.We developed a parallelizer Paractive to prove our thoughts, and tested Paractive using CFD programs from PDL. Parallel programs obtained can achieve 80% when executed on 4 computing nodes, and 70% on 8 nodes of YH-III super computer.