📄 http:^^www.cc.gatech.edu^fac^karsten.schwan^
字号:
Date: Wed, 20 Nov 1996 22:47:19 GMT
Server: Apache/1.1.1
Content-type: text/html
Content-length: 20848
Last-modified: Thu, 25 Jul 1996 14:54:49 GMT
<HEAD><TITLE>Karsten Schwan Home Page </TITLE></HEAD><BODY><P><P><H1> Karsten Schwan <BR> </h1>RESEARCH STATEMENT:<BR><B>Programming and Operating Systems for High Performance Platforms</B><BR> </b><P> College of Computing<BR>Georgia Institute of Technology<BR>Atlanta, GA 30332<BR>(404)894-2589, schwan@cc.gatech.edu<BR><P><P>Prof. Schwan has been working with experimental and commercial parallel machines since their inception in the late 70's. His researchis driven by relevant applications developed as part of a multidisciplinarylaboratory at Georgia Tech -- the HPPCEL Laboratory contains uniprocessorand multiprocessor supercomputers (a Cray machine, an IBM SP-2, andnetworked sets of workstations like IBM RS6000, SUN and SGI machines). Hisresearch is also driven by the future heterogeneous SMP and distributedsystems infrastructure experienced in commercial and research settings --the Distributed Laboratories project is constructing a heterogeneouscomputing environment consisting of ATM-networked uniprocessor and SMP machines,jointly offering the computational power and interconnection bandwidth ofmodern supercomputers.<P><B>Applications and the <!WA0><A HREF="http://www.cc.gatech.edu/net/www/systems/projects/HPPCEL/"> HPPCEL</A> Laboratory.</B>The HPPCEL Laboratory for experimentation with parallel applicationsis undertaking the development of several sample, large-scale parallel applications jointly with end users (<!WA1><A HREF="http://www.cc.gatech.edu/grads/e/Greg.Eisenhauer/papers/particle_paper.ps.Z>molecular dynamics, </A><A HREF="ftp://ftp.cc.gatech.edu/pub/coc/tech_reports/1995/GIT-CC-95-17.ps.Z">atmospheric modeling,</A>finite element codes, fluid flow applications). Ineach case, these `applications' not onlyconsist of specific computational tasks, but also include online analyses being performed for these computations and real-time visualizations of program data and monitoring of performance information, all of which are executed simultaneously such that end users can inspect program results and steerprogram executions to gain experimental insights. Additional applicationsused to evaluate operating system support, program adaptation, andprogramming environments research have included:(a) robotics applications and real-time simulations (the ASV autonomoussuspension vehicle, schema based robot navigation, and discreteevent simulations), (b) real-time multi-media applications involvingmultiple machines and human end users, (c) the design and implementation of interactive graphical interfaces for high performanceapplications, and (d) graph-based searching algorithms used in the solution ofOperations Research problems.<P><B><!WA2><A HREF="http://www.cc.gatech.edu/systems/projects/DISTLABS/">Distributed Laboratories.</A></B> The topic of our joint research is the realization of distributed laboratories, where individuals can interact with each other,and more importantly, with powerful, distributed computational tools asreadily as if all were located in a single site. Our intent is to permitscientists, engineers, and managers at geographically distinct locations(including individuals ``tele-commuting'' from home) to combine theirexpertise in solving shared problems, by allowing them tosimultaneously view, interact with, and steer sophisticatedcomputations executing on high performance distributed computing platforms.Prof. Schwan's primary contribution to this effort concerns the construction ofa distributed laboratory for experimentation with high performancenumeric computations for applications in molecular physics,atmospheric sciences, working with high performance atmospheric and pollution modelling, and manufacturing systems. In addition, Prof. Schwan'sgroup is involved in the development of basic technologies facilitating theuse of future heterogeneous and large-scale distributed and parallel systems.These technologies are concerned with ``middleware'' softwarethat will be directly utilized by distributed laboratory applications anddescribed in more detail below:<UL><LI>The Falcon project is studying the dynamic monitoring, adaptation, and interactive steering of high performance computations for on-line control of ``virtual laboratory instruments'' and for ``what-if?'' experimentation with complex simulation models by distributed laboratory users. <LI> The COBS project is concerned with the distributed systems andcommunication technologies underlying distributed laboratory applications. The emphasis in distributed systems research is on support for shared-statein multi-granular and distributed computing environments.Research in communication protocols is concerned with providing the necessarycommunication latencies, bandwidths, and transmission guarantees across heterogeneous underlying communication media.</UL><P>Prof. Schwan's current research addresses three basic research topicsaddressing the computational requirements and underlying hardware systems presented by the HPPCEL and Distributed Laboratory efforts: <UL><LI> programming support tools, <LI> operating systems, and<LI> communication protocols.</UL><P><B>Programming tools - FALCON.</B> The <!WA3><A HREF="http://www.cc.gatech.edu/systems/projects/FALCON/"> FALCON</A> project addresses the construction of interactive parallel programs, even permitting programmers to `steer'such applications. Steering may be defined as:<P> <B> Program Steering - </B>A program that can interact with human users during its execution,where the purpose of such interactions is the use of on-line programinformation to affect or `steer' program execution.<P>The FALCON project's goal is to have high performance programs become equally interactive as commonly used PC tools (e.g., spreadsheets) such thatprogram performance is either maintained or improved! The expectation isthat scientists can work more effectively when able to view program resultsduring execution and steer program execution toward useful data domains. Theinitial experience is that on-line steering can improve the performance ofprograms by avoiding needless execution and by adapting (algorithmically)specific program abstractions to better match their functionality to dynamically varying program characteristics. In essence,we expect to gain significant increases in the usefulness of parallelprograms to end users and in program performance by use of steering.<P>FALCON offers specific support for program performancemonitoring, evaluation, and tuning, and they provide system integrationsupport using an incore data dictionary. Using FALCON, information required for performance tuning is easily captured, analyzed, displayed, and then used for program performance improvement. Program monitoring is performed in an identical fashion for both shared and non-shared memory parallel machines, using language-based and graphical user interfaces. Using programming extensions of FALCON now under development, program tuning is performed graphically, as well, by explicit manipulation of parallel programs displayed as hierarchical sets of related objects. The tools will be used in the context of large-scale parallel applications beinginteractively steered by end users. The FALCON tool isoperational on SUNs, SGIs, IBM RS/6000, IBM SP-2 and the KSRmachine. FALCON's programming extensions are still under construction, but their runtime support is already operational.<P><B>Adaptable Operating Systems.</B>The efficient execution of a parallel programrequires that operating system primitives match the program's performanceand reliability requirements. By constructing and evaluating noveloperating system mechanisms for multiple parallel architectures,we are studying high-performance operating system support for multipleparallel applications, including real-time applications, scientific and engineering codes, discrete event simulations, and optimization codes.<P>This research has resulted in several software libraries that facilitatethe construction of parallel programs. The <!WA4><A HREF="http://www.cc.gatech.edu/systems/projects/Cthreads/">`Cthreads' </A>library is commonly used in operating systems and parallel computing courses and is available on many uniprocessor and multiprocessor machines. Its extensions toward `Configurable CThreads' address the on-linemonitoring and configuration of high performance parallel programs, ranging from the rapid adaptation of individual program abstractions to user-directed program steering. Specific examples of configurable abstractions constructed with the CThreads basis include (1) <!WA5><A HREF="ftp://ftp.cc.gatech.edu/pub/coc/tech_reports/1993/GIT-CC-93-05.ps.Z">adaptable multiprocessor locks</A>able to improve program performance by adapting lock characteristics to program behavior during program execution and (2) <!WA6><A HREF="ftp://ftp.cc.gatech.edu/pub/coc/tech_reports/1993/GIT-CC-93-22.ps.Z">adaptable communication protocols</A><P>The DSA (Distributed Shared Abstractions) library for construction of distributed objects is available on shared and (in a prototype version) on distributed memory machines, including networks of workstations. <! VERNARD, need the link to DSA here...software library available via remote FTP. Bother Bodhi if you cannot find it.>Its purpose is to support programmers in the construction of efficient program abstractions capable of spanning both shared memory and distributed memory platforms. Its initial SMP prototype is now being extended to addressboth sets of platforms, including larger scale distributed systems (see theCOBS project described below).<P>The <!WA7><A HREF="ftp://ftp.cc.gatech.edu/pub/coc/tech_reports/1994/GIT-CC-94-11.ps.Z">Kernel ToolKit (KTK) </A>is a substrate for parallel programming and for the development of multiprocessor operating system kernels. At theobject-level, it gives developers the ability to construct custom implementations of object-based operating system abstractions. Objects arecustomized using meta-object-like interfaces described by `object attributes'and implemented by `policies'. Since policies may be present at runtime,KTK objects may also be dynamically adapted to suit varying runtime requirementsof parallel and distributed applications. KTK`s attributes are the principalbasis for construction of adaptable objects in the COBS project describednext.<P>The <!WA8><A HREF="http://www.cc.gatech.edu/systems/projects/COBS>COBS </A>project's purposeis the development of technologies for high performance implementations ofobjects across SMP and networked platforms. Its initial prototype providesboth DSM and DSA support on networked systems, and it offers developersbasic mechanisms for programming diverse object abstractions, ranging from shared blocks of memory of varied consistency to typed objects offering well-defined operational interfaces, with varying methods for invoking both. The intent of COBS is to interoperate with COM objects while also permitting⌨️ 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 + -