Data Movement in DoAll Loops

Jackson, Daniel Thomas

Permalink

https://hdl.handle.net/2142/126866 Copy

Description

Title

Data Movement in DoAll Loops

Author(s)

Jackson, Daniel Thomas

Issue Date

1985-08

Keyword(s)

: DoAll Loops, DoAll Scheduling, Global and Local Memory Moves, Global Network, Horizontal Spreading, Multiprocessors, Prefetching Data, Simple Compiler Generated Testing, Vertical Spreading

Date of Ingest

2025-03-25T13:58:07-05:00

Abstract

: Our aim is to give program data movement statistics that can help parallel computer architects to design faster multiprocessors. We present the results from our study of 432 DoAll loops from scientific packages. From the study, we give meaningful statistics such as the number of global moves per operation in DoAll loops. From these results, we show how the prefetch/execute/poststore architecture might be changed to increase network utilization. We consider ways to schedule DoAll loops at compile time and discuss the advantages and disadvantages of each scheme. Finally, we show how to use these compile-time scheduling techniques to allow the sharing of data between DoAlls that were split up by mode floating and the sharing of data between DoAlls and Serial loops.

Publisher

Center for Supercomputing Research and Development, University of Illinois at Urbana-Champaign

Series/Report Name or Number

CSRD-524, UILU-ENG-85-8014

Type of Resource

text

Genre of Resource

technical report

Language

eng

Sponsor(s)/Grant Number(s)

National Science Foundation US NSF DCR84-10110; US NSF DCR84-06916; U.S. Department of Energy US DOE DE-FG02-85ER25001

Owning Collections