📄 http:^^www-net.cs.umass.edu^mcast.html
字号:
Date: Tue, 14 Jan 1997 19:04:34 GMT
Server: NCSA/1.5.1
Last-modified: Tue, 24 Sep 1996 02:28:27 GMT
Content-type: text/html
Content-length: 7069
<title>Research in Multicast Networking</title><body><!WA0><img src="http://images/Lines/rainbow.gif"> <p><H2> Research in Multicast Networking </H2> <!WA1><img src="http://images/Lines/rainbow.gif"> <p>The recent success of multicast applications such as Internet teleconferencing, distributed interactive simulation, anddata dissemination applications illustrate the tremendous potential of applications built upon wide-area multicast communicationservices. Our ongoing research is focussed intwo crucial areas of multicast networks: reliable multicast protocolsand call admission in multicast networks.<h3> Reliable Multicast Protocols </h3>While some multicast applications (such as video and voice) do not require reliable data transfer,others (such as shared whiteboards and data dissemination) do. The requirement of reliable data transfer for this last set of applications posesa difficult challenge to network designers - howto design and implement a reliable multicast protocol that can handle 100s or 1000sof participants.We are currently pursuing several related efforts in the area of reliable multicast protocols:<ul><li>A critical issue for multicast applications and thehigher layer protocols that support them is the manner in which packetlosses occur within the multicast network.Thus, one goal of our recent work <!WA2><a href="ftp://gaia.cs.umass.edu/pub/Yajn96:Loss.ps.Z"> [Yajnik et al., 1996] </a>has been to examinethe <i> spatial and temporal correlation in packet loss</i> amongparticipants in a multicast session. (Informally, by ``spatially'' correlatedloss, we mean the loss, i.e., lack of reception, of the same packet at many sites; by ``temporally'' correlated loss, wemean the loss of consecutive packets at a given receiver.)<p> In <!WA3><a href="ftp://gaia.cs.umass.edu/pub/Yajn96:Loss.ps.Z"> [Yajnik et al., 1996] </a>we present and analyze packet loss data collectedsimultaneously at up to 12 hosts at geographically distinct locations in Europe and the US.These hosts are connected via the MulticastBackbone (MBone) network Our results show that<ul><li>For most of the traces, the loss on the backbone links of the MBone multicast network is observed to be small (2% or less), as compared to the average loss seen by a receiver. However, due to occasional outages lasting from few seconds to few minutes, in some backbone links, the spatially correlated loss between receivers does go up to 20%, in a few datasets.<li> There is a significant amount of burst loss (consecutive losses) at each site. One or more extremely long loss bursts, lasting from a few seconds up to 3 minutes (around 2000 consecutive packets), occur in almost every trace. <li> Most of the loss bursts consist of isolated single losses, but the few very long loss bursts contribute heavily to the total packet loss.<li> Some receivers see periodic packet loss lasting for approximately 0.6sec. (8 consecutive packets) and occurring at 30 sec. intervals.</ul><p><li>A second effort in the area of reliable multicast <!WA4><a href="ftp://gaia.cs.umass.edu/pub/Tows96:Comparison.ps"> [Towsley et al., 1996] </a>examines two different approaches to providingreliable, scalable multicast communication. The <i>sender-initiated</i> approachplaces the responsibility for providing reliable multicaston the sender, which maintains stateinformation on all receivers to which it is multicasting.This is accomplished by having receivers return positive acknowledgments (ACKs)for every correctly received packet, and having the sender use timers to detect potential packet losses. The alternateapproach, a <i> receiver-initiated </i> approach, shifts most of the responsibility for reliable data deliveryto the receivers. Each receiver is responsible for detecting lostpackets and informing the sender via negative acknowledgments (NAKs)when it requires the retransmission of a packet. <p>In the case of an application consisting of a single sender transmitting reliably to many receivers (referred to as a<i> one-many </i> application) we observe throughsimple analyses that a simple receiver-initiated protocol whichrequires receivers to return negative acknowledgments (NAKs) to thesender over point-to-point channels provides substantially betterperformance (in terms of the maximum supportable throughputof successfully transmitted messages)than a sender-initiated protocol. Further substantialimprovement is obtained by the multicasting of NAKs coupled with theintroduction of random delays prior to the transmission of a NAK.In the case of an application where all participants act as both sendersand receivers (referred to as a <i> many-many </i> application), simpleanalyses illustrate that the receiver-initiated protocol whichrequires receivers to return negative acknowledgments (NAKs) to thesender over point-to-point channels almost doubles throughput overa sender-initiated counterpart. However, unlike the one-many case, the multicastingof NAKs in the many-many scenariodoes not fare as well, leading to only a small increase in throughput over the sender-initiated counterpart.<p><li>A more recent work of ours, done jointly withProfessor Miki Yamamoto of Osaka University,builds upon this previous workby presenting adelay analysis of the three generic sender- and receiver-initiated protocols identified in our earlier work.Our results indicate that no protocol has uniformly better delaybehavior than the others.At low packet arrival rates and moderate to high loss rates, we find that the ACK-based protocol has a significantly smaller delay than either of the NAK-based approaches. We also find thatthe depending on the network loss rates, either of twoNAK based protocols (once which delivers NAKs point-to-point to thesender, and one which multicast NAKs to receivers)has better performance at high arrival rates.</ul>In addition to these research efforts, we also have ongoingwork in the areas of multicast flow control, and the development and analysis of an approachs towards relaible multicast that use multiple multicast groupsfor error recovery purposes<h3> Call Admission in Multicast Networks </h3>(This part still under construction).<p><P><b> References</b><BR><P><ul><LI> M. Yajnik, J. Kurose, D. Towsley, ``Packet Loss Correlation in the MBone Multicast Network,'' to appear in <i>IEEE Global Internet Conf.</i> (London, Nov. 1996). <!WA5><a href="ftp://gaia.cs.umass.edu/pub/Yajn96:Loss.ps.Z"> [postscript] </a><LI> D. Towsley, J. Kurose, S. Pingali, ``A Comparison of Sender-Initiated and Receiver-Initiated Reliable Multicast Protocols, to appear <i>IEEE Journal on Selected Areas in Communications.</i> <!WA6><a href="ftp://gaia.cs.umass.edu/pub/Tows96:Comparison.ps"> [postscript] </a></ul></body><P><ADDRESS><I>kurose@gaia.cs.umass.edu <BR>Tue Sep 10 20:30:18 EDT 1996</I></ADDRESS>⌨️ 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 + -