📄 http:^^www.cs.rochester.edu^u^leblanc^prediction.html
字号:
Date: Wednesday, 20-Nov-96 20:04:49 GMTServer: NCSA/1.3MIME-version: 1.0Content-type: text/htmlLast-modified: Friday, 21-Jun-96 21:14:58 GMTContent-length: 8223<HTML><HEAD><TITLE>Performance Prediction</TITLE></HEAD><BODY BACKGROUND="textures/yellow_paper.gif"><!WA0><IMG Align=middle SRC=http://www.cs.rochester.edu/u/leblanc/icons/carnival.gif><H2>Measurement and Prediction of <BR>Parallel Program Performance</H2><H3>NSF CISE grant CCR-9510173</H3><H3>Faculty</H3><DL><DT><!WA1><A HREF="http://www.cs.rochester.edu/users/faculty/leblanc/">Tom LeBlanc</A><EM>(leblanc@cs.rochester.edu)</EM></DT><DD></DD><DT>in collaboration with<!WA2><A HREF="http://cs-www.bu.edu/faculty/crovella/Home.html">Mark Crovella</A><EM>(crovella@cs.bu.edu)</EM><DT><DD></DD></DL><H3>Graduate Students</H3><DL><DT><!WA3><A HREF="http://www.cs.rochester.edu/users/grads/meira/">Wagner Meira</A><EM>(meira@cs.rochester.edu)</EM></DT><DD></DD><DT><!WA4><A HREF="http://www.cs.rochester.edu/users/grads/poulos/">Alex Poulos</A><EM>(poulos@cs.rochester.edu)</EM></DT><DD></DD><DT><!WA5><A HREF="http://www.cs.rochester.edu/users/grads/nikolaos/">Nikolaos Hardavellas</A> <EM>(nikolaos@cs.rochester.edu)</EM></DT><DD></DD></DL><HR><H3>Project Summary</H3><P>Carnival is a tool designed to automate the processof understanding the performance of parallel programs.It supports performance measurement, modeling, tuning,and visualization.<P>Carnival measurements are based on<em>predicate profiling</em>,which quantifies the time spent in each category of overhead during execution.Our first implementation of predicate profilingwas implemented on the KSR-1.We now have implementations forthe SGI Challenge multiprocessor,the IBM SP-2,a network of SUN workstations running PVM,and a cluster of Alpha workstations.<P>Carnival is a novel attempt to automate the cause-and-effectinference process for performance phenomena.In particular, Carnival currently supports <em>waiting time analysis</em>,an automatic inference process that explains each sourceof waiting time in terms of the underlying causes,instead of simply identifying where it occurs.We are now developing a similar technique to explainthe causes of communication.<P>Our ultimate goal is to combine the accuracy of empirical performancemeasurement with the predictive power of analytic performance modeling.Towards that end, Carnival supports<em>lost cycles analysis</em>,which uses a priori knowledge of the sources and characteristicsof the overhead categories in parallel systems to guide and constrainthe modeling process.The Lost Cycles Toolkit, which we are integrating within Carnival,combines empirical model-building techniques from statisticswith measurement and modeling techniques for parallel programs.<P>Carnival is also a visualization toolthat provides a link between performance measurementsand the source code.The interface presents the original source code in a window.Along the left hand side of the source is a grey-scalescroll bar that indicates the amount of time spent ineach portion of the source code (summed across all processors).Along the right hand side of the source code are color bars that indicatethe percent of time spent in each overhead category by that section ofsource code (again summed across all processors).Pop-up windows are used during modeling and waiting time analysis.<P>The Carnival implementation comprises about 15,000lines of Tcl/Tk and C source code.It has been installed at the Cornell Theory Centerand we plan to make it more widely available soon.We are currently porting the instrumentation library(the only machine-dependent portion of the tool)to clusters of DEC Alphas connected by the DEC Memory Channel.<HR><H3>Related Publications</H3><UL><LI> Crovella, M. R. Bianchini, T. LeBlanc, E. Markatos, and R. Wisniewski, ``Using Communication-to-Computation Ratio in Parallel Program Design and Performance Prediction,'' In Proc. 4th IEEE Symposium on Parallel and Distributed Processing, Dallas, Texas, December 1992, pp. 238-245.<LI> Crovella, M. and T. LeBlanc. <!WA6><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/93.WPDD.performance_predicates.ps.Z">Performance Debugging Using Parallel Performance Predicates.</A> In Proc., 3rd ACM/ONR Workshop on Parallel and Distributed Debugging, pp. 140-150, May 1993.<LI> Crovella, M. and T. J. LeBlanc. <!WA7><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/93.tr479.the_search_for_lost_cycles.ps.Z">The Search for Lost Cycles: A New Approach to Performance Tuning of Parallel Programs.</A> In Proc. Supercomputing '94, Washington, D.C. November 1994, pp. 600-609. Also available as Technical Report 479, Computer Science Dept., Univ. of Rochester, December 1993.<LI> Crowl, L. A., M. Crovella, T. J. LeBlanc, and M. L. Scott. ``The Advantages of Multiple Parallelizations in Combinatorial Search.'' J. of Parallel and Distributed Computing, 21, 1, April 1994, pp. 110-123. Also available as <!WA8><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/93.tr451.beyond_data_parallelism.ps.Z">Technical Report 451</A>, Computer Science Dept., Univ. of Rochester, April 1993.<LI> Crovella, M. <!WA9><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/95.tr573.Performance_prediction_and_tuning_of_parallel_programs.ps.gz">Performance Prediction and Tuning of Parallel Programs.</A> Ph.D. Dissertation, TR 573, Computer Science Department, University of Rochester, August 1994.<LI> Mark Crovella, Thomas J. LeBlanc, Wagner Meira, Jr., <!WA10><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/95.tr580.Performance_measurement_and_modeling_with_lost_cycles_toolkit.ps.gz">Parallel Performance Prediction Using the Lost Cycles Toolkit.</A> TR 580, Computer Science Department, University of Rochester, May 1995.<LI> Wagner Meira Jr., <!WA11><A HREF="ftp://ftp.cs.rochester.edu/pub/papers/systems/95.tr589.Modeling_performance_of_parallel_programs.ps.gz">Modeling Performance of Parallel Programs</A>, TR 589, Computer Science Department, University of Rochester, June 1995.<LI> Wagner Meira Jr., Thomas J. LeBlanc, and Alexandros Poulos, <!WA12><A HREF="http://www.cs.rochester.edu/u/leblanc/papers/carnival/paper.html">Performance Visualization and Tuning with Carnival</A>, ACM SIGMETRICS Symp. on Parallel and Distributed Tools, May 1996, pp. 1-10.</UL><HR><H3>Related Projects</H3><UL> <LI> <!WA13><A HREF="http://www.nas.nasa.gov/NAS/Tools/Projects/AIMS/">AIMS</A> (NASA Ames) <LI> <!WA14><A HREF="http://www.cc.gatech.edu/systems/projects/FALCON/falcon_home.html>Falcon</A> (Georgia Tech) <LI> <A HREF="http://www-pablo.cs.uiuc.edu/">Pablo</A> (University of Illinois) <LI> <!WA15><A HREF="http://www.cs.wisc.edu/~paradyn">Paradyn</A> (University of Wisconsin) <LI> <!WA16><A HREF="http://www.cs.uoregon.edu/~bhelm/poirot/index.html">Poirot</A> (University of Oregon) <LI> <!WA17><A HREF="http://www.cc.gatech.edu/computing/Architecture/projects/tass.html">TASS</A> (Georgia Tech) <LI> <!WA18><A HREF="http://www.cs.uoregon.edu:80/paracomp/tau/">TAU</A> (University of Oregon)</UL><HR><H3>Other Information</H3><P>There is a new symposium on parallel and distributed toolssponsored by ACM SIGMETRICS.The symposium has a<!WA19><A HREF="http://www.cs.wisc.edu/~paradyn/Call.html">home page</A>,and was held May 1996 as part of the ACM Federated Conferencein Philadelphia, PA.See the proceedings of that symposium for papers on the latest workin this area.<P>The Parallel Tools Consortium was formed to help coordinate tooldevelopment in the parallel processing community.See their <!WA20><A HREF="http://www.llnl.gov/ptools/ptools.html">home page</A>for a list of projects approved by the consortium, as well as acomprehensive list of research projects on parallel tools.<P>The Cornell Theory Center maintains a<!WA21><A HREF="http://www.tc.cornell.edu/UserDoc/Software/PTools/">listof parallel tools</A>(including Forge, ParaGraph, ParaScope, and upshot)and associated documentation.<P>For an introduction to parallel computing, and a discussion ofrelated performance issues, see Ian Foster's on-line text<!WA22><A HREF="http://www.mcs.anl.gov/dbpp">Designing and BuildingParallel Programs</A>.<HR><ADDRESS>last modified June 21 1996 / Tom LeBlanc / leblanc@cs.rochester.edu</ADDRESS></BODY></HTML>⌨️ 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 + -