Cypress CSC-1200T Uživatelská příručka stáhnout pdf (Strana 62)

62 Cray T3E User’s Guide

Chapter 7

Interprocess communication

This chapter describes how to use the MPI or PVM message-passing li-

braries on the Cray T3E at CSC. In addition, the properties of the Cray

data-passing library SHMEM are described in some detail. The data-

parallel High Performance Fortran (HPF) programming model is intro-

duced, too.

7.1 The communication overhead

Parallelization on the T3E can be done by three diﬀerent approaches:

message"-passing, data-passing and data-parallel programming. The

deﬁning feature of message passing is that the data transfer from the

local memory of one process to the local memory of another requires op-

erations to be performed by both processes. In contrast, data-passing

routines for sending or receiving data are one-sided. In data-parallel

model, the programmer speciﬁes only the data distribution between

processes, and the compiler generates a parallel executable code.

The fastest way to communicate between processors on the Cray T3E is

using the SHMEM library (Shared Memory Library). This is a Cray-speciﬁc

data-passing library and will not produce portable code to other archi-

tectures. It may be preferable to use MPI (Message Passing Interface),

a standardized and portable message-passing library. Another portable

message-passing library on the T3E is the somewhat older PVM (Parallel

Virtual Machine), which is, in general, a bit slower than MPI. The Cray T3E

system has also a data-parallel HPF (High Performance Fortran) compiler,

supporting Cray’s CRAFT data-parallel programming model, as well.

Message latency (or start-up time) is about 1 µs (microseconds) when

using the SHMEM library. With MPI the latency is about 30 µs. The

maximum bandwidth is in practice about 230 MB/s with both SHMEM

1 2 ... 57 58 59 60 61 62 63 64 65 66 67 ... 123 124

Žádné komentáře

Cypress CSC-1200T Uživatelská příručka Strana 62