📄 node34.html

📁 Design and building parallel program
💻 HTML
字号:
<html><!DOCTYPE HTML PUBLIC "-//W3O//DTD W3 HTML 2.0//EN">
<!Converted with LaTeX2HTML 95.1 (Fri Jan 20 1995) by Nikos Drakos (nikos@cbl.leeds.ac.uk), CBLU, University of Leeds >
<HEAD>
<TITLE>3.8 Input/Output</TITLE>
</HEAD>
<BODY>
<meta name="description" value="3.8 Input/Output">
<meta name="keywords" value="book">
<meta name="resource-type" value="document">
<meta name="distribution" value="global">
<P>
 <BR> <HR><a href="msgs0.htm#2" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#2"><img ALIGN=MIDDLE src="asm_color_tiny.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/asm_color_tiny.gif" alt="[DBPP]"></a>    <A NAME=tex2html2261 HREF="node33.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node33.html"><IMG ALIGN=MIDDLE ALT="previous" SRC="previous_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/previous_motif.gif"></A> <A NAME=tex2html2269 HREF="msgs0.htm#24" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#24"><IMG ALIGN=MIDDLE ALT="next" SRC="next_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/next_motif.gif"></A> <A NAME=tex2html2267 HREF="node26.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node26.html"><IMG ALIGN=MIDDLE ALT="up" SRC="up_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/up_motif.gif"></A> <A NAME=tex2html2271 HREF="node1.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node1.html"><IMG ALIGN=MIDDLE ALT="contents" SRC="contents_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/contents_motif.gif"></A> <A NAME=tex2html2272 HREF="node133.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node133.html"><IMG ALIGN=MIDDLE ALT="index" SRC="index_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/index_motif.gif"></A> <a href="msgs0.htm#3" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#3"><img ALIGN=MIDDLE src="search_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/search_motif.gif" alt="[Search]"></a>   <BR>
<B> Next:</B> <A NAME=tex2html2270 HREF="msgs0.htm#24" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#24">3.9 Case Study: Shortest-Path Algorithms</A>
<B>Up:</B> <A NAME=tex2html2268 HREF="node26.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node26.html">3 A Quantitative Basis for Design</A>
<B> Previous:</B> <A NAME=tex2html2262 HREF="node33.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node33.html">3.7 A Refined Communication Cost Model</A>
<BR><HR><P>
<H1><A NAME=SECTION02480000000000000000>3.8 Input/Output</A></H1>
<P>
An important determinant of performance in many parallel programs is
<A NAME=3971>&#160;</A>
the time required to move data between memory and secondary storage,
<A NAME=3972>&#160;</A>
that is, the time required for input/output (I/O).  Applications with
substantial I/O requirements include the following.
<P>
<UL><LI>
<em> Checkpoints</em>.  Many computations performed on parallel computers
<A NAME=3975>&#160;</A>
execute for extended periods.  Periodic checkpointing of computation
state is essential in order to reduce the cost of system failures.  On
large parallel computers, state can be large (many gigabytes).
<P>
<LI>
<em> Simulation data.</em> Scientific and engineering simulations that
compute the time evolution of physical systems periodically save
system state for subsequent visualization or analysis.  Some
simulations can generate very large amounts of data---hundreds of gigabytes
or more in a single run.
<P>
<LI>
<em> Out-of-core computation.</em> Some programs must operate on data
<A NAME=3978>&#160;</A>
structures that are larger than available ``core'' memory.  In
principle, a virtual memory system can perform the necessary paging of
data to and from disk.  In practice, not all parallel computers
provide virtual memory. Even when they do, application performance can
often be enhanced by explicit management of data movement.
<P>
<LI>
<em> Data analysis.</em> Many applications involve the analysis of large
amounts of data.  For example, climate model output or data from
weather satellites may be searched for ``extreme events'' such as high
temperature values, a video database may be searched for specified
images, and a database of credit card transactions may be searched for
patterns of fraudulent use.  These data analysis applications are
particularly demanding from an I/O point of view, because relatively
little computation is performed on each datum retrieved from disk.
</UL>
<P>
It is difficult to provide a general discussion of parallel I/O
because different parallel computers have radically different I/O
architectures and hence parallel I/O mechanisms.  However, we can make
several points that have wide applicability.
<P>
We can often gain a reasonable understanding of the cost of an I/O
operation by thinking of it as a communication from the processors
<A NAME=3981>&#160;</A>
that perform that operation to one or more disks.  The cost of a disk
I/O operation can be approximated by a startup cost and a per-word
transfer cost in much the same way as an interprocessor
communication.  (However, the startup cost is typically much greater.)
As in interprocessor communication, the keys to good performance are
to maximize the utilization of available paths and to minimize
startups.
<P>
<P><A NAME=5177>&#160;</A><IMG BORDER=0 ALIGN=BOTTOM ALT="" SRC="img590.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/img590.gif">
<BR><STRONG>Figure 3.22:</STRONG> <em> I/O architecture of an idealized parallel computer.  
<em> P</em>
 processors are connected by multiple I/O channels to
<em> D</em>
 disks.</em><A NAME=figioarch>&#160;</A><BR>
<P>
<P>
If a computer has only a single disk or if multiple disks are
connected to a single processor, little can be done to
optimize I/O performance.  However, in practice most parallel
computers provide multiple paths from processors to disk, whether by
providing distinct ``I/O nodes'' or by connecting disks directly to
processors (Figures <A HREF="node34.html#figioarch" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node34.html#figioarch">3.22</A>).  On architectures of this
sort, we seek to organize I/O operations so that multiple processors
read and write concurrently, using multiple paths.  Thus, centralized
I/O strategies that cause data to pass through a single processor are
unlikely to be efficient and are certainly not scalable.
<P>
In addition to maximizing concurrency in I/O operations, we need to be
<A NAME=3988>&#160;</A>
concerned about the number of distinct read or write requests required
to transfer data between disk and processor memory.  This can often
have a greater impact on I/O performance than can the amount of data
transferred.  The number of I/O requests depends partly on how data
are distributed on disk and in memory.  The distribution in memory will
presumably be determined by the application; the distribution on disk
either will be under programmer control or will be selected by the
file system.  Data may sometimes be ``striped'' (scattered) across
available disks so as to reduce the probability of many processors
attempting to read from the same disk simultaneously.
<P>
If distributions on disk and in memory differ, then a large number of
reads or writes may be required in order to achieve data transfer.
This problem is analogous to what happens when transferring data
structures between two parallel program components that require
different distributions.  As will be discussed in
<A NAME=3989>&#160;</A>
Chapter <A HREF="node39.html#chapmod" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node39.html#chapmod">4</A>, at least two approaches are possible in this
situation: we can modify one or both components to use different
distributions, or we can explicitly redistribute data before or while
transferring it.  Because I/O requests tend to be more expensive than
interprocessor communications, it is often better to perform an
explicit redistribution of data in memory so as to minimize the number
of I/O requests.  This leads to a two-phase access strategy, in which
the data distributions used on disk and in memory are decoupled.  The
merits of these various approaches can be explored analytically with
performance models.
<P>

<BR> <HR><a href="msgs0.htm#2" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#2"><img ALIGN=MIDDLE src="asm_color_tiny.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/asm_color_tiny.gif" alt="[DBPP]"></a>    <A NAME=tex2html2261 HREF="node33.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node33.html"><IMG ALIGN=MIDDLE ALT="previous" SRC="previous_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/previous_motif.gif"></A> <A NAME=tex2html2269 HREF="msgs0.htm#24" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#24"><IMG ALIGN=MIDDLE ALT="next" SRC="next_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/next_motif.gif"></A> <A NAME=tex2html2267 HREF="node26.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node26.html"><IMG ALIGN=MIDDLE ALT="up" SRC="up_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/up_motif.gif"></A> <A NAME=tex2html2271 HREF="node1.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node1.html"><IMG ALIGN=MIDDLE ALT="contents" SRC="contents_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/contents_motif.gif"></A> <A NAME=tex2html2272 HREF="node133.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node133.html"><IMG ALIGN=MIDDLE ALT="index" SRC="index_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/index_motif.gif"></A> <a href="msgs0.htm#3" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#3"><img ALIGN=MIDDLE src="search_motif.gif" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/search_motif.gif" alt="[Search]"></a>   <BR>
<B> Next:</B> <A NAME=tex2html2270 HREF="msgs0.htm#24" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#24">3.9 Case Study: Shortest-Path Algorithms</A>
<B>Up:</B> <A NAME=tex2html2268 HREF="node26.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node26.html">3 A Quantitative Basis for Design</A>
<B> Previous:</B> <A NAME=tex2html2262 HREF="node33.html" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/node33.html">3.7 A Refined Communication Cost Model</A>
<BR><HR><P>
<P><ADDRESS>
<I>&#169 Copyright 1995 by <A href="msgs0.htm#6" tppabs="http://www.dit.hcmut.edu.vn/books/system/par_anl/tppmsgs/msgs0.htm#6">Ian Foster</a></I>
</ADDRESS>
</BODY>
⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -