Next: Computer Terminology
Up: Introduction
Previous: Background and Objectives
Computer benchmarks are computer programs that form standard tests of the
performance of a computer and the software through which it is used. They
are written to a particular programming model and implemented by specific
software, which is the final arbiter as to what the programming model is.
PARKBENCH
has initially adopted two such models:
- Fortran77 + PVM: This is the classical distributed-memory MIMD
model in which a number of separate logical processors execute
asynchronously independent Fortran77 programs in their individual
and separate memory space. The only communication and synchronisation
between these programs is by sending messages containing data using
the PVM (Parallel Virtual Machine [2]) library of Fortran
communication subroutines.
- High Performance Fortran (HPF): This is an extension of the
classical SIMD model in which a single instruction stream in the
Fortran90 language [4] specifies operations that apply,
notionally simultaneously, to vectors and higher-order arrays of data.
In HPF [5] data distribution statements are added by the
programmer as comments to the Fortran90 program to help the compiler
generate efficient code on a distributed-memory computer system.
A benchmark is therefore testing a software interface to a computer, and
not a particular type of computer architecture. For example, benchmarks
using the "F77+PVM" programming model can be run on any computer providing
this interface, both distributed-memory message-passing computers which
have message-passing hardware, and shared-memory computers which lack the
hardware but can simulate message-passing in software.
top500@rz.uni-mannheim.de
Tue Nov 14 15:43:14 PST 1995