An Efficient Reservation Connection Control Protocol for Gigabit Networks

An Efficient Reservation Connection Control Protocol for Gigabit Networks
of 22
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  Ž . Computer Networks and ISDN Systems 30 1998 1135–1156 An efficient reservation connection control protocol for gigabitnetworks  1 Emmanouel A. Varvarigos  ) , Vishal Sharma  Data Transmission and Networking Laboratory, Department of Electrical and Computer Engineering, Uni Õ ersity of California, Santa Barbara, CA 93106-9560, USA Abstract Ž . The  Efficient Reser  Õ ation Virtual Circuit   protocol or ERVC is a novel connection control protocol designed forconstant-rate, delay-insensitive traffic in gigabit networks. We explain the operation of the protocol, discuss its features andadvantages, and present its performance characteristics. The ERVC protocol is appropriate for sessions that require anexplicit reservation of capacity and can tolerate the round-trip delay associated with the reservations. In the ERVC protocol,the durations of the sessions are recorded, and every node keeps track of the  utilization profile  of each outgoing link, whichdescribes the amount of residual capacity available on the link as a function of time. This feature allows capacity to bereserved only for the duration of the session, starting at the time it is actually needed. Therefore, the protocol utilizescapacity considerably more efficiently than regular reservation schemes do and results in markedly lower blockingprobability for new sessions. The ERVC protocol also has the ‘‘reservation ahead’’ feature, which allows a node to calculatethe time at which the requested capacity will be available and reserve it in advance, avoiding in this way the wastefulrepetition of the call setup phase.  q 1998 Elsevier Science B.V. All rights reserved. Keywords:  Constant-rate traffic; Connection establishment; Protocol; Reservation-based; Performance characteristics 1. Introduction In the era of gigabit networking, which is rapidlybecoming a reality with advances in VLSI technol-ogy and fiber-optic transmission systems, networkswill be limited by the propagation delay on the w x channel 1 . This is a limitation imposed by physicsthat will not change with improved implementation.The important network issues of connection estab-lishment, flow control, buffering, and congestioncontrol therefore need to be re-examined in this ) Corresponding author. 1 Research supported by ARPA under Contract DABT63-93-C-0039. changed communication environment. As Partridge w x 2 in his series of thought questions on the chal-lenges of gigabit networking rightly points out, highbandwidth-delay product and increased propagationlatency are two crucial factors that will differentiategigabit networks from most present day networks,and that will impact not only the performance of such networks, but also the protocols designed tomanage them and the applications designed to usethem.Propelled by the advances in new communicationtechnologies, a number of prototype high-speed net-works have been developed in the US and Europe todate. These include the PARIS high-speed network  Ž  w x  . see 3,4 for various aspects of its design , and its 0169-7552 r 98 r $19.00 q  1998 Elsevier Science B.V. All rights reserved. Ž . PII: S0169-7552 98 00009-9  ( ) E.A. Var  Õ arigos, V. Sharma r Computer Networks and ISDN Systems 30 1998 1135–1156  1136 successor the plaNET r ORBIT high-speed network  w x 5 that will be used in the AURORA gigabit testbed w x 6 . Aurora is one of the five gigabit testbeds beingoverseen by the Corporation for National Research w x Initiatives 7 . The other networks in this initiative w x w x w x include the Vistanet 8 , CASA 9 , Blanca 10 , and w x Nectar 11 gigabit testbeds. Other gigabit researchinitiatives within the US include the MIT Lincon w x Labs.-based all-optical network consortium 12 .High-speed networks outside the US include theMultiG project funded by the Swedish Institute of  w x Computer Science 13 , the Cambridge backbone w x ring in Britain 14,15 , the Berkom project in Ger- w x w x many 16 , and TEN-34 17 , a new gigabit network testbed deployed in Europe.The ERVC protocol, which is the subject of thispaper, is part of the connection and flow controlprotocols designed for the Thunder and Lightning Ž  w x . network see, for instance, 18–21 , which is cur-rently being built at UCSB under the sponsorship of  w x DARPA 22 . The Thunder and Lightning network isa virtual-circuit switched, fiber-optic network thatwill operate at serial link speeds of up to 40 Gbit r s,and is projected to carry a diverse mix of traffictypes. Our objectives in designing the connectionand flow control algorithms for this network are toensure lossless transmission, efficient utilization of capacity, minimum pre-transmission delay fordelay-sensitive traffic, and packet arrival in correctorder. The two new connection control protocols thatwe have proposed to meet these objectives are the Ž . Ready-to-Go Virtual Circuit or RGVC protocol and Ž . the Efficient Reservation Virtual Circuit or ERVC w x protocol. The RGVC protocol 20 , which will beused for best-effort service and for traffic that haslittle tolerance for delay, uses back-pressure andbuffering at intermediate nodes, whereas the ERVCprotocol, which is appropriate for constant-rate ses-sions that require guaranteed bandwidth, uses ex-plicit reservations and requires little buffering atintermediate nodes.The very high link speeds in the Thunder andLightning network were proposed because of therealization that at least a sizable portion of traffic infuture gigabit networks would involve high-speedtransfer of massive amounts of data at nearly con-stant rates, and would require guaranteed losslessdelivery and an explicit reservation of bandwidth.The ERVC protocol, proposed in this paper, providesefficient transfer of such data, and has been moti-vated by several design considerations. First, forseveral applications such as high-speed program-datatransfer between supercomputers or bulk file trans-fer, loss-based traffic integration will not be practical w x 23 . This is because providing a reliable transportservice requires that, ideally, not even a single cell Ž be lost here loss refers to the loss due to statistical . multiplexing and not that due to transmission errors ,because each such loss can force retransmission of large volumes of data. Clearly, in gigabit networkslike the Thunder and Lightning network, the band-width-delay product, being very large, can result inthe discarding of substantial amounts of data in caseof retransmissions if efficient lossless transfer is notprovided. Second, for high-speed file-transfer typeapplications, long burst transmissions can easilyoverload the network, unless they have pre-negoti-ated at least a minimum bandwidth with the network.This has also been realized by several other re-searchers, many of whom have advocated burst-basedbandwidth reservation as a viable and prudent choice Ž  w x w x see Hiu 24 , Ohnishi et al. 25 , Suzuki and Tobagi w x w x . 26 , and Iwata et al. 27 . Therefore, from the pointof view of both transmission integrity and network efficiency, traffic of this type should be transferredonly after a specific and explicit allocation of re-sources precedes each data burst.The ERVC protocol is employed for call setup if the session is not critically delay-sensitive and re-quires an explicit reservation of bandwidth. In whatfollows, the word session will be applied both to anew session, or to a new burst for an ongoing sessionthat has a virtual path to the destination but mayhave no bandwidth reserved on it before hand. Newsessions are generated at each source with a speci-fied destination, duration, tolerable delay, and band-width requirement. For a new session, a path withadequate residual capacity is computed at the source,based on the topology and link utilization informa-tion that it has at that time. For a burst belonging to acontinuing session, this computation is performed atthe connection set up phase and is not required atthis time. In the ERVC protocol, the duration of a Ž . session or burst is recorded during the call estab-lishment phase, and each network node keeps track of the capacity available on its outgoing links as a  ( ) E.A. Var  Õ arigos, V. Sharma r Computer Networks and ISDN Systems 30 1998 1135–1156   1137 function of time. A setup packet is sent over the pathto make the appropriate reservations and set therouting tables. Each intermediate node reserves therequired capacity starting at the time at which this Ž capacity will actually be used which is at least oneround-trip delay after the arrival of the setup packet . at the node , and for a length of time equal to thesession duration. If the session duration is unknown,it is treated as infinite, and capacity is reserved for Ž that session for an unspecified duration as in regular . reservation schemes . If the capacity is not availableat the requested time, the setup packet may make areservation starting at the first time the capacitybecomes available. Therefore, the ERVC protocolhas the ‘‘reservation ahead’’ feature, in the sensethat capacity may be reserved in advance for use at alater time. This differs from other reservation virtualcircuit protocols, where session durations are notrecorded and capacity is reserved starting at the timethe setup packet arrives at a node. The idea of advance reservation of resources has been used inthe past mainly in connection with video-telecon- Ž  w x . ferencing centers see 28–31 . In these works, there Ž . is a single-server or multi-server central reservationoffice that users call to request conferencing facili-ties for a given future date. This is quite differentthan the reservation ahead feature of the ERVCprotocol, where the objective is the efficient utiliza-tion of the network links, and capacity is reserved ina distributed manner, through the collaboration of nodes on a session’s path.In the ERVC protocol, capacity is blocked forother sessions only for the duration of the call, whatwe call the  timed reser  Õ ation  feature of the protocol,and is available for the remaining time. As thediscussion of Section 2 and the performance resultspresented in Section 4 will indicate, this is particu-larly important for high-speed networks, where prop-agation times are large compared to transmissiontimes, because it reduces blocking and allows agreater number of sessions to be served. It alsoavoids the wasteful repetition of the call establish-ment process, because it enables a session to reservethe required capacity in its first attempt, probably ata time later than the requested time. If the time atwhich adequate bandwidth first becomes available isunacceptable because of the delay requirements of the session, the call is blocked and is reattemptedlater, probably using a different path. The ERVCprotocol requires a pre-transmission delay at leastequal to the round-trip propagation delay between Ž the source and the destination as all reservation . protocols do .Several implementation issues, some of which wediscuss in the latter sections of our paper, are impor-tant for the correct and efficient operation of theERVC protocol. Correct timing is important in orderto ensure that data transmission starts after all reser-vations are made and terminates before any interme-diate node releases the reserved capacity. Also, sincereservations are made in a distributed, asynchronousway, through the collaboration of the nodes on asession’s path, and capacity on a link may not beavailable at the time requested for by the setuppacket but at some future time, it is important toensure that the time intervals reserved on each link are consistent with each other. For the ERVC proto-col to function efficiently, the queueing and process-ing delays of control packets have to be small andpredictable, and the uncertainties and round-off er-rors in recording the various time parameters have tobe controlled. The information required by the proto- Ž . col rates and session durations has to be recordedand processed in an efficient way. As we discuss inSection 3, we use a linked-list structure to store therates and durations of the sessions, and record timesas relative times, which enables a fast list update tobe performed at a switch. Finally, the protocol has tocope with link and node failures.The remainder of the paper is organized as fol-lows. In Section 2 we provide the motivation for theERVC protocol by discussing its advantages overother reservation protocols for high-speed networks,and we describe some applications that would bene-fit from it. In Section 3 we explain the operation of the protocol. Specifically, in Section 3.1 we describethe call setup procedure, and in Section 3.2 wedescribe the data structures that are required and theway they are updated. The connection control actionsperformed by the source, intermediate, and destina-tion nodes are outlined in Section 3.3. In Section 3.4we comment on the delays experienced by controlpackets, and discuss how the protocol handles uncer-tainties in timing. In Section 4, we present theperformance characteristics of the ERVC protocol,and we compare its performance to that of regular  ( ) E.A. Var  Õ arigos, V. Sharma r Computer Networks and ISDN Systems 30 1998 1135–1156  1138 reservation schemes. Finally, in Section 5 we givesome concluding remarks, while in the appendiceswe give the details of the connection control actionsperformed at the nodes and discuss how the protocolcopes with link and node failures. 2. Why the ERVC protocol? In this section, we discuss some drawbacks of other reservation schemes and explain how the ERVCprotocol helps to overcome them. We use the term Ž immediate reservation virtual circuit schemes ab- . breviated IRVC to refer to schemes where the ca-pacity required by a session at an intermediate nodeis reserved starting at the time the setup packetarrives at the node, and which, unlike the ERVCprotocol, do not allow for reservations to be made Ž for future time instants we discuss this aspect . shortly . This includes several recently proposedreservation schemes such as the FRP r DT protocol Ž Fast Reservation Protocol with Delayed Transmis- .  w x sion proposed by Boyer and Tranchier 32 , the fast w x bandwidth reservation schemes of Suzuki et al. 26 , Ž . the fast resource management FRM protocols men- w x tioned by Fotedar et al. 33 and discussed in detail w x by Tranchier et al. 34 , and the connection establish- w x ment scheme proposed by Cidon et al. 3 . Thescheme proposed by Cidon et al. uses a logical treeto execute a distributed capacity check algorithm thatspeeds up the reservation phase, but it still suffers Ž . even though to a lesser extent from a drawback common to immediate reservation schemes as wenow explain. A cause of the inefficiency in theseschemes arises because the capacity reserved for thesession is actually used at least one round-trip delayafter the arrival of the setup packet at the node. Thisis because the setup packet has to travel from theintermediate node to the destination, an acknowl-edgement has to be sent from the destination to thesource, and the first data packet of the session has to Ž travel from the source to the intermediate node see . Fig. 1 .Over long transmission distances, the round-trippropagation delay may be comparable to, or evenlarger than, the holding time of a session. In particu-lar, if a typical session requests capacity  r   bits r s,and transfers a total of   M   bits over a distance of   L Fig. 1. IRVC, ERVC, and RGVC protocols for the case where the setup packet is successful in making the appropriate reservations. In Ž . IRVC protocols, where session durations are not recorded, the capacity is blocked for duration equal to  M  r r   q T   , where  T   is the rt rt  roundtrip propagation delay. In the ERVC protocol, capacity is blocked for other sessions only for the holding time  M  r r  . In the RGVC Ž . protocol a tell-and-go type of protocol that will also be used in the Thunder and Lightning network , capacity is occupied for time  M  r r  w x plus the time offset between the transmission of the setup packet and the first data packet of the session. In the RGVC protocol 20 , thesetup packet is first transmitted along the path, followed after a short offset-interval by the data packets, with back-pressure exercised if needed.  ( ) E.A. Var  Õ arigos, V. Sharma r Computer Networks and ISDN Systems 30 1998 1135–1156   1139 kilometers, then the maximum percentage of timethat the capacity is efficiently used in an IRVCprotocol is  M  r r e s  , 1 Ž . 2  L n  r c q  M  r r  where  n  r c s 5  m s r km is the propagation delay inthe fiber. Typical values of the above parameters forgigabit networks like the Thunder and Lightningnetwork are expected to be  r  s 10 Gbit r s,  M  s 0.2 Ž Gbit corresponding, for instance, to the transfer of amedium-sized file in a high-speed data-transfer oper- . Ž . ation , and  L s 1500 km long haul communication ,which yields  e s 0.57. In other words, for the aboveparameters, the capacity reserved by a typical sessionwith an IRVC protocol stays idle for a round-tripdelay of 15 ms, and is used for time equal to thetypical holding time of a session, which is equal to20 ms. This efficiency factor  e  becomes even smalleras  r   or  L  increase, or  M   decreases. In contrast, theefficiency factor  e  for the ERVC protocol can be aslarge as  e s 1, independently of the parameters  r  ,  L ,and  M  . In the simulations results presented in Sec-tion 4, we show that the ERVC protocol can achievelink utilization close to one while keeping the block-ing probability for new sessions below some reason-able threshold, say 0.1. By contrast, IRVC protocols Ž achieve their maximum utilization which is always . less than that of the ERVC protocol at the cost of considerable blocking of new sessions, which im-poses a further penalty on the network through mul-tiple reservation attempts.To understand the advantage of recording sessiondurations and the ‘‘reservation ahead’’ feature of theERVC protocol, consider the situation shown in Fig.2 where a setup packet requests 10 Gbit r s of capac-ity on an outgoing link   l  that has only 5 Gbit r s of capacity available at that time. If an IRVC protocolis used, such a call will be blocked. Since, however,reservations on link   l  are such that 10 Gbit r s of capacity will become available after 14 ms, and thefirst data packets of the new session will arrive at the Ž . link after at least 30 ms a round-trip delay , thesession should be accepted. If the ERVC protocol isused, such a call will be accepted, because each noderecords the session durations and the setup packetwill be able to reserve 10 Gbit r s of capacity starting Fig. 2. The advantage of recording session durations in the ERVCprotocol and its ‘‘reservation ahead’’ feature. at a time 30 ms after its arrival at the node. TheERVC protocol also has the ‘‘reservation ahead’’feature, which allows sessions to reserve capacityahead in time and avoids repetition of the call setupphase. To see this, assume that the round-trip delayof the setup packet in Fig. 2 is only 10 ms, while theacceptable delay of the session is 18 ms. The setuppacket now requires 10 Gbit r s of capacity on link   l starting at a time 10 ms after its arrival at theintermediate node. Since 10 Gbit r s of capacity isavailable on link   l  after 14 ms and this time is withinthe range of delays that can be tolerated by thesession, the call will be accepted on its first attempt.Thus, the reservation ahead feature avoids unneces-sarily prolonged call setup phases and the associatedwaste of bandwidth, reduces a session’s susceptibil-ity to blocking, and leads to efficient utilization of the available capacity. We discuss the performanceadvantages that result from this feature when pre-senting our simulation results in Section 4.We see at the present time, several potentialapplications where the ERVC protocol will be use-ful. One such application is in the area of distributed w x network computing 35 , especially computing in-volving several powerful supercomputers that willneed to exchange bulk data, relatively frequently andwithout any loss. Virtual circuit connections with aminimum guaranteed bandwidth will have to be setup between these sites to facilitate the efficient trans- w x fer of data when needed 36 . For geographicallyseparated sites where the round-trip delay is compa-rable to or larger than the data transmission time, itwill be most efficient to reserve capacity only for the
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks