http:^^www.cs.arizona.edu^people^pete^
来自「This data set contains WWW-pages collect」· EDU^PEOPLE^PETE^ 代码 · 共 185 行
EDU^PEOPLE^PETE^
185 行
Date: Thu, 21 Nov 1996 19:52:34 GMTServer: NCSA/1.5Content-type: text/htmlLast-modified: Tue, 19 Nov 1996 18:49:38 GMTContent-length: 5274<head><title>Peter J. Downey's Home Page</title></head><body><a name="pagetop"><hr><h1> <!WA0><img align=bottom src="http://www.cs.arizona.edu/people/pete/pete2.gif"> Peter J. Downey </h1><hr>Peter J. Downey is Associate Professor of <!WA1><a href=http://www.cs.arizona.edu>Computer Science</a>at the <!WA2><a href=http://www.arizona.edu>University of Arizona</a>.He received a Ph.D. from Harvard University in 1974, and joinedthe faculty at the Pennsylvania State University.In 1978 he moved to Arizona.<p>Downey's main research area is probabilistic analysis ofalgorithms and systems, with particular emphasis on modelingthe performance of parallelalgorithms.These areas are described in detail below.<p>During the 1996-97 academic year, Downey is on sabbaticalleave, but can be reached at the addresses below.<hr><address>Department of Computer Science <br>Gould-Simpson Building, Room 739<br>The University of Arizona <br>PO BOX 210077<br>Tucson AZ 85721-0077<br>Phone: (520) 621-2207 <br>FAX: (520) 621-4246 <br><!WA3><a href=mailto:pete@cs.arizona.edu><tt>pete@cs.arizona.edu</tt></a></address><hr><p><BR>Naval Observatory Time is now: <!WA4><IMG SRC="http://tycho.usno.navy.mil/cgi-bin/nph-usnoclock.gif?zone=MMT&ticks=04" ALT="Clock"> <P><p><a name = papers"><H2>Parallel Program Performance</H2>The goal of this research project isthe development ofanalytic methods to evaluate the performance of parallel asynchronousalgorithms and systems.In order to understand the origins and effects of overheads in parallelcomputation, such as synchronization delays,methods focus on probabilistic models of workloads andtask processing times.Rather than developing exact but unrealistic models with only narrowapplicability, emphasis is on the development of bounds and approximationsthat are robust and valid for extremely general workload and taskdistributions (distribution-free methods).For large workloads and massive numbers of parallel processors, laws oflarge number apply and allow run-time and speed-up to be approximated orbounded in terms of a few underlying parameters of the workload.<H3>Analysis of Overheads</H3>Analysis of parallelalgorithms is distinctly different from that ofsequential algorithms because of the presence of overheads,following from the need to coordinate parallel activity. Overheadsare of two fundamentally different kinds.<I>Explicit</I>overheadsresult from the need to add additional codeto a parallel algorithm to handle coordination matters suchas forks and message sends.<I>Implicit</I>overheadsarise from delays spent waiting on synchronizationevents, such as joins,to occur.Implicit and explicit overheadsincrease with increasing degree of parallelism;expected implicit overheadsincrease with the mean, variance and skewness of the underlying task timedistributions.Much of this research aimsto quantify the effectsof such overheads,and to understand how they depend onvarious parameters of the workload distribution.Work on this subject includes<li><!WA5><a href="http://www.cs.arizona.edu/people/pete/papers/bounds.ps">Bounds and Approximations for Overheads in the Time to JoinParallel Forks </a>,<i> ORSA Journal on Computing 7</i>, 2(Spring 1995), 125-139.<H3>Foundations: Extreme Order Statistics</H3>Let <I>X1, ... , Xn</I>be randomvariables, representing execution times of tasks.Two random variables derived from the task timesplay a criticalrole in determining the performance ofparallel activity:<UL><LI>the extreme<I>Mn = max ( X1, ... , Xn)</I>and<LI>the sum<I>Sn = X1 + ... + Xn.</I></UL>Work aimed atunderstanding thebehavior of both<I>Sn</I>and<I>Mn</I> for large<I>n</I>,even ifthetask times are dependent, forms a fundamental partof this research upon which all analysis is based.This foundationaltheory must meet a very practical methodological constraint:to be able to bound or approximatethe behavior ofextrema and sumswithout detailed information about the task workload distributions.Work in this area includes<li><!WA6><a href="http://www.cs.arizona.edu/people/pete/papers/ratio.ps"> The Ratio of the Extreme to the Sumin a Random Sequence with Applications</a>,with P.E. Wright,University of Arizona Technical Report TR 94-18 (1994).<H3>Scheduling and Sequencing</H3>Parallel performance issues occur at the operating systems levelin the design of multiprocessor schedulers.The design and analysis of simplescheduling policies having guaranteedperformance bounds is an important part of this research.New applications of stochastic ordering relations to proveoptimality of static scheduling policies andto develop bounds for dynamic policies are beingextended as part of this work.New work in this area includes<li><!WA7><a href="http://www.cs.arizona.edu/people/pete/papers/scheduling.ps">Scheduling Independent Tasksto Minimize the Makespan on Identical Machines </a>,with E.G. Coffman, Jr. and J. Bruno,<i>Probability in the Engineering and InformationalSciences 9</i>, 3(Fall 1995), 447-456.<p><!WA8><img src="http://www.cs.arizona.edu/images/hrule.gif" height=3 width=1000 alt="------------------------------------------------------------------"><br>[ <!WA9><a href="#pagetop">Top of Page</a>|<!WA10><a href="http://www.cs.arizona.edu/"">Department Home Page</a> ]<small><br>http://www.cs.arizona.edu/people/pete/<br>Last updated July 11, 1996</small></body>⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?