A DISTRIBUTED PROGRAMMING MODEL AND ITS APPLICATIONS TO COMP

Recent advances in high performance computing architectures have presented a clear trend that future systems must include computers from different classes. It is conceivable that effective large scale computing in general must be done in a heterogenous dis

A DISTRIBUTED PROGRAMMING MODEL AND ITS

APPLICATIONS TO

COMPUTATION INTENSIVE PROBLEMS FOR

HETEROGENEOUS ENVIRONMENTS

Yuan Shi

Department of Computer and Information Sciences

Temple University

Philadelphia, PA 19122

shi@fac.cis.temple.edu

(215)787-6437

(Published in AIP Conference Proceedings 283, Earth and Space Science InformationSystems, Pasadena, CA 1992, Editor: Arthur Zygielbaum, pp. 827-848)

ABSTRACT

Recent advances in high performance computing architectures have presented a clear trendthat future systems must include computers from different classes. It is conceivable thateffective large scale computing in general must be done in a heterogenous distributedenvironment. The reported research seeks to build a generic virtual processor model on topof heterogeneous computing and communication devices. By using a specification basedapproach, we can effectively customize the available heterogeneous devices for everycomputer application. In this paper, we shall present the computational results of three fieldapplications in scientific visualization, engineering simulation and financial simulation usinga Scatter-And-Gather method (or virtual vector processing). To aid objective evaluation ofthe virtual processor model, we also include the program re-engineering costs for achievingsuch performances.

Keywords: Distributed Heterogeneous Computing, Distributed Operating System.

1. INTRODUCTION

High computing efficiency can be achieved by parallelizing an application over a givencomputing architecture. In this article we intend to generalize the commonly knownapproaches to parallelize an application over a set of heterogeneous computingarchitectures by organizing coarse grain parallel components.

There are three basic types of parallelizable components in every computing application:SIMD, MIMD and pipelined. In order to exploit the full potential of the availablecomputing powers and existing parallelism of a given application, the granularity of parallelcomponents must vary to optimally offset the communication latency. If the distributed

Recent advances in high performance computing architectures have presented a clear trend that future systems must include computers from different classes. It is conceivable that effective large scale computing in general must be done in a heterogenous dis

environment is non-volatile, i.e. it is single user oriented, a parallel compiler can producefairly optimized codes for a given hardware architecture. Heterogeneous software andcommunication protocols in typical volatile distributed environments have made bothbuilding a generic distributed operating system and a "heterogeneous parallel compiler"very difficult.

The main focus of the reported research is to promote a virtual processor model that canbe constructed dynamically on top of heterogeneous computers and a software system(SYNERGY, 1990 U.S. Patent pending) to make such a model feasible. In particular, weshall report the computation and re-engineering results using coarse grain SIMDcomponents for three field applications.

2. THE VIRTUAL PROCESSOR MODEL

The proposed virtual processor model consists of only three types of components: virtualSIMD, virtual MIMD and virtual pipeline. A processor assignment with respect to anexecution environment for a computing application defines the virtual processor for thatapplication. Obviously there is a virtual processor defined for every currently runningdistributed or non-distributed application.

The central idea of this virtual processor model is to customize a set of distributedcomputers for a given application by fitting a network of virtual SIMD, MIMD andpipelined components to an application's natural dataflow structure. Numerous tuningdevices must be constructed to counter react to unexpected situations in typical volatileenvironments.

The interface of the virtual processor consists of a databus network of a given applicationand a distributable program -> processor mapping. The databus network is a network ofdistributable programs interconnected through databuses. Each distributable program is anindependent process that can be dynamically loaded onto a range of processors. Eachdatabus is a user defined abstract (distributed) data object for which a set of pre-definedoperations can be applied. For example, a generic queue (or mailbox) can be defined as adatabus along with its operations: open, close, read, write and post. A tuple space can alsobe a databus along with operations: open, close, put, read and get. A small set of suchobjects is suffice for most scientific computing problems.

The databuses are the essential media for building virtual parallel components. Forexample, the use of two or more tuple space objects can be used to construct a virtualvector processor and virtual pipe must employ a series of generic queues. Techniques usedin vectorizing compilers to discover vectorizable elements can be applied here to discovercoarse grain vectors with minor modifications.

A raw sequential program cannot be readily executed on such a virtual processor, if goodperformance is expected. A re-engineering process must be carried out to relax the internaldataflows of the given program. This will be further discussed in the programming example

Recent advances in high performance computing architectures have presented a clear trend that future systems must include computers from different classes. It is conceivable that effective large scale computing in general must be done in a heterogenous dis

section.< …… 此处隐藏：26570字，全部文档内容请下载后查看。喜欢就下载吧 ……