?? http:^^www.cs.umd.edu^projects^harmony^
字號:
Date: Wed, 20 Nov 1996 19:40:16 GMT
Server: Apache-SSL/0.4.3b
Content-type: text/html
Content-length: 4715
Last-modified: Mon, 11 Nov 1996 21:09:25 GMT
<HTML><HEAD><TITLE>Active Harmony</TITLE></HEAD><BODY BACKGROUND="http://www.cs.umd.edu/~keleher/bkgnd.jpg"><center><h1>Active Harmony</h1><br><!WA0><IMG SRC="http://www.cs.umd.edu/~keleher/new.gif"><!WA1><a href="#discuss">Discussion Group</a> - <!WA2><a href="http://www.cs.umd.edu/projects/harmony/related.html">Links</a></center><!WA3><IMG SRC="http://www.cs.umd.edu/Images/Maryland_Bar_Top.gif"><BODY> Complex software systems today can be characterized by distribution,heterogeneity, and changing resource requirements and capacities.These attributes make static systems unsuitable for a wide range of tasksthat need high performance, or are long-lived.In order to achieve high performance in such environments for more than ashort period of time, systems need to dynamically adapt to changingresource capacities and application requirements.We are designing and building <em>Active Harmony</em>, a softwarearchitecture that supports distributed execution of computational objectsin such environments through the following innovations:<ul><li><b>dynamic execution environment:</b> Dynamic adaptation to network andresource capacities, both when computational objects are created, and whenapplication requirements or resource capacities change.Active Harmony will attempt to maximize data affinity and load balancingthrough intelligent resource allocation and object migration.<li><b>automatic application adaptation:</b> A framework that permits runtime adaptation of algorithms, data distribution, and loadbalancing. Active Harmony will export a detailed metric interface toapplications, allowing them to access processor, network, and operatingsystem parameters.Applications export tuning options to the system, which can then automaticallyoptimize resource allocation.Measurement and tuning can therefore become first class objects in theprogramming model.Programmers can write applications that include ways to adapt computation toobserved performance and changing conditions.<li><b>shared-data interfaces:</b> Active Harmony will support shared-memorysemantics among computational objects regardless of location, allowing bothpeer-to-peer and client-server computations to exploit the simplifiedprogramming model and fine-grained sharing permitted by traditionalshared-memory environments.Innovations include support for heterogeneity of both data \emph{and}program code, a multi-level security scheme that adapts data and codeinterfaces to the degree of trust between computational objects, andsupport for the dynamic execution environment.</ul>The unique aspect of the Active Harmony work is the emphasis on adapting toheterogeneous and changing environments.Other researchers have studied some of the constituent issues that we planto address.Our emphasis is on the inter-relationships <em>between</em> objects in thesystem.The primary result of this research will be an infrastructure and a set ofalgorithms that permit global resource optimization under changingconditions.<!WA4><IMG SRC="http://www.cs.umd.edu/Images/Maryland_Bar_Top.gif"><h3>Project Members:</h3><ul><li><!WA5><a href="http://www.cs.umd.edu/~keleher">Dr. Jeff Hollingsworth</a><li><!WA6><a href="http://www.cs.umd.edu/~keleher">Dr. Pete Keleher</a></ul><BR CLEAR=RIGHT><br><p><!WA7><IMG SRC="http://www.cs.umd.edu/Images/Maryland_Bar_Top.gif"><a name="discuss"><h3>Discussion group:</h3>We will be forming a reading and discussion group to investigate the aboveissues. Check back later (week of 11/18) for details.<p><b>Papers</b> (partial list):<ul><LI><I><!WA8><A HREF="http://http.cs.berkeley.edu/~harchol/lb.ps">ExploitingProcess Lifetime Distributions for Dynamic Load Balancing</A></I></LI><LI><I><!WA9><A HREF="http://HTTP.CS.Berkeley.EDU/~dusseau/Papers/sigmetrics96.ps.gz">EffectiveDistributed Scheduling of Parallel Workloads</A></I></LI><LI><I><!WA10><A HREF="http://HTTP.CS.Berkeley.EDU/~dusseau/Papers/sigmetrics.ps">The Interaction of Parallel and Sequential Workloads on a Network of Workstations</A></I></LI><li><i><!WA11><a href="http://cs-tr.cs.washington.edu:80/Dienst/Repository/2.0/Body/ncstrl.uwash_cse%2fTR-95-10-01/postscript">Using Runtime Measured Workload Characteristics in Parallel Processor Scheduling</i></a></UL><p><b>Open questions:</b><ul><li>Is the fastest sequential process migration appropriate for parallelapplications? i.e. should more process state be moved prior tore-scheduling a parallel application?<li>How can scheduling of parallel and client/server applications take dataaffinity into account?<li>Is co-scheduling as important if we don't care about latency, onlythroughput?<li>How does co-scheduling perform when we move from local networks to internets?</ul><!WA12><IMG SRC="http://www.cs.umd.edu/Images/Maryland_Bar_Top.gif"><I>Last updated on November 11, 1996</I></BODY>?? 快捷鍵說明
復制代碼
Ctrl + C
搜索代碼
Ctrl + F
全屏模式
F11
切換主題
Ctrl + Shift + D
顯示快捷鍵
?
增大字號
Ctrl + =
減小字號
Ctrl + -