1

Published in IEEE/ACM Transactions on Networking, February 1997.

A Measurement-based Admission Control Algorithmfor Integrated Services Packet Networks (Extended Version)

Sugih Jamin, Peter B. Danzig, Scott J. Shenker, and Lixia Zhang

Abstract— Many designs for integrated servicesnetworks offer a bounded delay packet delivery ser-vice to support real-time applications. To providebounded delay service, networks must use admissioncontrol to regulate their load. Previous work on ad-mission control mainly focused on algorithms thatcompute the worst case theoretical queueing delayto guarantee an absolute delay bound for all pack-ets. In this paper we describe a measurement-basedadmission control algorithm for predictive service,which allows occasional delay violations. We havetested our algorithm through simulations on a widevariety of network topologies and driven with vari-ous source models, including some that exhibit long-range dependence, both in themselves and in theiraggregation. Our simulation results suggest thatmeasurement-based approach combined with the re-laxed service commitment of predictive service en-ables us to achieve a high level of network utilizationwhile still reliably meeting delay bound.

I. BOUNDED DELAY SERVICES ANDPREDICTIVE SERVICE

There have been many proposals for supportingreal-time applications in packet networks by pro-

Sugih Jamin was supported in part by the Uniforum Re-search Award and by the Office of Naval Research Labora-tory under contract N00173-94-P-1205. At USC, this researchis supported by AFOSR award number F49620-93-1-0082, bythe NSF small-scale infrastructure grant, award number CDA-9216321, and by equipment loan from Sun Microsystems, Inc.At PARC, this research was supported in part by the AdvancedResearch Projects Agency, monitored by Fort Huachuca undercontract DABT63-94-C-0073. The views expressed here donot reflect the position or policy of the U.S. government.

Sugih Jamin and Peter B. Danzig are with the Com-puter Science Department, University of Southern California,Los Angeles, California 90089-0781 (email: jamin@usc.edu,danzig@usc.edu).

Scott J. Shenker is with the Xerox Palo Alto Re-search Center, Palo Alto, California 94304-1314 (email:shenker@parc.xerox.com).

Lixia Zhang is with the Computer Science Department, Uni-versity of California at Los Angeles, Los Angeles, California90095 (email: lixia@cs.ucla.edu).

viding some form of bounded delay packet deliv-ery service. When a flow requests real-time ser-vice, it must characterize its traffic so that the net-work can make its admission control decision. Typ-ically, sources are described by either peak and av-erage rates [FV90] or a filter like a token bucket[OON88]; these descriptions provide upper boundson the traffic that can be generated by the source.The traditional real-time service provides a hardor absolute bound on the delay of every packet;in [FV90], [CSZ92], this service model is calledguaranteed service. Admission control algorithmsfor guaranteed service use the a priori characteriza-tions of sources to calculate the worst-case behaviorof all the existing flows in addition to the incomingone. Network utilization under this model is usu-ally acceptable when flows are smooth; when flowsare bursty, however, guaranteed service inevitablyresults in low utilization [ZF94].

Higher network utilization can be achieved byweakening the reliability of the delay bound. Forinstance, the probabilistic service described in[ZK94] does not provide for the worst-case sce-nario, instead it guarantees a bound on the rate oflost/late packets based on statistical characteriza-tion of traffic. In this approach, each flow is allottedan effective bandwidth that is larger than its averagerate but less than its peak rate. In most cases theequivalent bandwidth is computed based on a sta-tistical model [Hui88], [SS91] or on a fluid flow ap-proximation [GAN91], [Kel91]) of traffic.1 If onecan precisely characterize traffic a priori, this ap-proach will increase network utilization. However,we think it will be quite difficult, if not impossi-ble, to provide accurate and tight statistical modelsfor each individual flow. For instance, the averagebit rate produced by a given codec in a teleconfer-

�

We refer the interested readers to [Jam95] for a more com-prehensive overview and bibliography of admission control al-gorithms.

ence will depend on the participant’s body move-ments, which can’t possibly be predicted in advancewith any degree of accuracy. Therefore the a prioritraffic characterizations handed to admission con-trol will inevitably be fairly loose upper bounds.

Many real-time applications, such as vat, nv,and vic, have recently been developed for packet-switched networks. These applications adapt toactual packet delays and are thus rather tolerantof occasional delay bound violations; they do notneed an absolutely reliable bound. For these toler-ant applications, references [CSZ92], [SCZ93] pro-posed predictive service, which offers a fairly, butnot absolutely, reliable bound on packet deliverytimes. The ability to occasionally incur delay vi-olations gives admission control a great deal moreflexibility, and is the chief advantage of predictiveservice. The measurement based admission con-trol approach advocated in [CSZ92], [JSZC92] usesthe a priori source characterizations only for in-coming flows (and those very recently admitted);it uses measurements to characterize those flowsthat have been in place for a reasonable duration.Therefore, network utilization does not suffer sig-nificantly if the traffic descriptions are not tight.Because it relies on measurements, and source be-havior is not static in general, the measurementbased approach to admission control can never pro-vide the completely reliable delay bounds neededfor guaranteed, or even probabilistic, service; thus,measurement-based approaches to admission con-trol can only be used in the context of predic-tive service and other more relaxed service com-mitments. Furthermore, when there are only afew flows present, the unpredictability of individ-ual flow’s behavior dictates that these measurementbased approaches must be very conservative—byusing some worst-case calculation for example.Thus a measurement based admission control algo-rithm can deliver significant gain in utilization onlywhen there is a high degree of statistical multiplex-ing.

In summary, predictive service differs in two im-portant ways from traditional guaranteed service:(1) the service commitment is somewhat less reli-able, (2) while sources are characterized by tokenbucket filters at admission time, the behavior of ex-isting flows is determined by measurement rather

than by a priori characterizations. It is importantto keep these two differences distinct because whilethe first is commonplace, the second, i.e. the useof measurement-based admission control, is morenovel. On the reliability of service commitment,we note that the definition of predictive service it-self does not specify an acceptable level of delayviolations. This is for two reasons. First, it is notparticularly meaningful to specify a failure rate to aflow with a short duration [NK92]. Second, reliablyensuring that the failure rate never exceeds a partic-ular level leads to the same worst-case calculationsthat predictive service was designed to avoid. In-stead, the CSZ approach [CSZ92] proposes that thelevel of reliability be a contractual matter betweena network provider and its customers—not some-thing specified on a per-flow basis. We presume thatthese contracts would only specify the level of vio-lations over some macroscopic time scale (e.g. daysor weeks) rather than over a few hundred packettimes.2 In this paper we describe a measurementbased admission control algorithm for predictiveservice. We demonstrate affirmative answers to thefollowing two questions. First, can one provide re-liable delay bounds with a measurement-based ad-mission control algorithm? Second, if one doesindeed achieve reliable delay bounds, does offer-ing predictive service increase network utilization?Earlier versions of this work have been published asreferences [JSZC92], [JDSZ95]. Incidentally, thework reported in this paper has been extended in[DKPS95] to support advance reservations. The au-thors of [DKPS95] have also replicated some of ourresults on their independently developed networksimulator.

The authors of [HLP93], [GKK95] use measure-ments to determine admission control, but the ad-mission decisions are pre-computed based on theassumption that all sources are exactly described byone of a finite set of source models. This approachis clearly not applicable to a large and heteroge-neous application base, and is very different fromour approach to admission control that is based

�

A network provider might promise to give its customerstheir money back if the violations exceed some level over theduration of their flow, no matter how short the flow; howeverwe contend that the provider cannot realistically assure that ex-cessive violations will never occur.

2

on ongoing measurements. In references [SS91],[AS94] the authors use measurement to learn theparameters of certain assumed traffic distributions.The authors of [DJM97], [Flo96] use measure-ment of existing traffic in their calculation of equiv-alent bandwidth, providing load, but not delay,bound. In references [Hir91], [CLG95], a neuralnetwork is used for dynamic bandwidth allocation.In [LCH95], the authors use pre-computed low fre-quency of flows to renegotiate bandwidth alloca-tion. Hardware implementation of measurementmechanisms are studied in [C

�91], [WCKG94].

II. MEASUREMENT-BASED ADMISSIONCONTROL FOR ISPN

Our admission control algorithm consists of twologically distinct aspects. The first aspect is theset of criteria controlling whether to admit a newflow; these are based on an approximate model oftraffic flows and use measured quantities as inputs.The second aspect is the measurement process it-self, which we will describe in Section III. In thissection we present the analytical underpinnings ofour admission control criteria.

Sources requesting service must characterize theworst-case behavior of their flow. In [CSZ92] thischaracterization is done with a token bucket filter.A token bucket filter for a flow has two parame-ters: its token generation rate, � , and the depth ofits bucket,

�. Each token represents a single bit;

sending a packet consumes as many tokens as thereare bits in the packet. Without loss of generality, inthis paper we assume packets are of fixed size andthat each token is worth a packet; sending a packetconsumes one token. A flow is said to conform toits token bucket filter if no packet arrives when thetoken bucket is empty. When the flow is idle ortransmitting at a lower rate, tokens are accumulatedup to

�tokens. Thus flows that have been idle for a

sufficiently long period of time can dump a wholebucket full of data back to back. Many non-constantbit rate sources do not naturally conform to a to-ken bucket filter with token rate less than their peakrates. It is conceivable that future real-time applica-tions will have a module that can, over time, learn asuitable � and

�to bound their traffic.

We have studied the behavior of our admissioncontrol algorithm mostly under the CSZ scheduling

discipline [CSZ92]. Under the CSZ scheduling dis-cipline, a switch can support multiple levels of pre-dictive service, with per-level delay bounds that areorder of magnitude different from each other. Theadmission control algorithm at each switch enforcesthe queueing delay bound at that switch. We leavethe satisfaction of end-to-end delay requirements tothe end systems. We also assume the existence ofa reservation protocol which the end systems coulduse to communicate their resource requirements tothe network.

When admitting a new flow, not only must the ad-mission control algorithm decide whether the flowcan get the service requested, but it must also decideif admitting the flow will prevent the network fromkeeping its prior commitments. Let us assume, forthe moment, that admission control cannot allowany delay violations. Then, the admission controlalgorithm must analyze the worst-case impact of thenewly arriving flow on existing flows’ queueing de-lay. However, with bursty sources, where the tokenbucket parameters are very conservative estimatesof the average traffic, delays rarely approach theseworst-case bounds. To achieve a fairly reliablebound that is less conservative, we approximate themaximal delay of predictive flows by replacing theworst-case parameters in the analytical models withmeasured quantities. We call this approximation theequivalent token bucket filter. This approximationyields a series of expressions for the expected max-imal delay that would result from the admission of anew flow. In CSZ, switches serve guaranteed trafficwith the weighted fair queueing scheduling disci-pline (WFQ) and serve different classes of predictivetraffic with priority queueing. Hence, the computa-tion of worst-case queueing delay is different forguaranteed and predictive services. In this section,we will first look at the worst-case delay computa-tion of predictive service, then that of guaranteedservice. Following the worst-case delay computa-tions, we present the equivalent token bucket filter.We close this section by presenting details of theadmission control algorithm based on the equiva-lent token bucket filter approximations.

A. Worst-case Delay: Predictive Service

To compute the effect of a new flow on exist-ing predictive traffic, we first need a model for

3

the worst-case delay of priority queues. Cruz, in[Cru91], derived a tight bound for the worst-casedelay, ���� , of priority queue level � . Our deriva-tion follows Parekh’s [Par92], which is a simpler,but looser, bound for � �� that assumes small packetsizes, i.e. the transmission time of each packet issufficiently small (as compared to other delays)and hence can be ignored. This assumption ofsmall packet sizes further allows us to ignore de-lays caused by the lack of preemption. Further, weassume that the aggregate rate, aggregated over alltraffic classes, is within the link capacity ( � � ���� ).

Theorem 1: Parekh [Par92]: The worst-caseclass j delay, with FIFO discipline within the classand assuming infinite peak rates for the sources, is

��� � � ������� ���� � ����������� � (1)

for each class j. Further, this delay is achieved fora strict priority service discipline under which classj has the least priority.3

The theorem says that the delay bound for class� is the one-time delay burst that accrues if theaggregate bucket of all classes � through � flowsare simultaneously dumped into the switch and allclasses � through ����� sources continue to send attheir reserved rates.

We now use Eq. 1 as the base equation to modelthe effect of admitting a new flow on existing pre-dictive traffic. First we approximate the traffic fromall flows belonging to a predictive class � as a singleflow conforming to a !#" �%$ � �'& token bucket filter. Aconservative value for " � would be the aggregate re-served rate of all flows belonging to class � . Next,we recognize that there are three instances when thecomputed worst-case delay of a predictive class canchange: (1) when a flow of the same class is ad-mitted, (2) when a flow of a higher priority class isadmitted, and (3) when a guaranteed flow is admit-ted. The switch priority scheduling isolates higherpriority ( (*) ) classes from a new flow of class ) ,so their worst-case delay need not be re-evaluatedwhen admitting a flow of class ) . In the remain-der of this section, we compute each of the three ef-fects on predictive traffic individually. At the end of+

For a proof of Theorem 1, we refer interested readers to[Par92], Theorem 2.4 or [Jam95], Theorem 1.

these computations, we will observe that admittinga higher priority predictive flow “does more harm”to lower priority predictive traffic than admitting ei-ther a guaranteed flow or a predictive flow of thesame priority.

In the equations below, we denote newly com-puted delay bound by � �-, . We denote the sumof guaranteed flows’ reservation by "/. . The linkbandwidth available for serving predictive traffic isthe nominal link bandwidth minus those reservedby guaranteed flows: � �0"1. .1. Effect of new predictive flow on same prioritytraffic.. We can model the effect of admitting a newflow of predictive class ) by changing the class’stoken bucket parameters to !#"1243 �%52 $ � 263 � 52 & , where! �%52 $ � 52 & are the token bucket parameters of the newflow:

879 � 9 ���:���;� �

���=@� � 9 ������� < ��A

� 9 A ��B9������ � 9 ������� < �

9CA ��B9���=@� � 9 ������� < �ED (2)

We see that the delay of class ) grows by a termthat is proportional to flow ’s bucket size.2. Effect of predictive flow on lower priority traf-fic.. We compute the new delay bound for class� , where � is greater than the requested class, ) , di-rectly from Eq. 1, adding in the bucket depth

� 52 andreserved rate � 52 of flow . 87� � 9 �

������F� � A � 9 A ��B9 A � � ��� 9HG �I� ����=@� � 9 ������� < � �=< 9 � �JB9 � � ���

���� 9HG � < �

� ���=@� � ���

��:�K� < ��L�=@� � ��������� < � � �JB9 A��B9���=< > � � ���

������ < � � �JB9NM O�P@QSRUTVM(3)where W is the number of predictive classes. The

first term reflects a squeezing of the pipe, in thatthe additional bandwidth required by the new flowreduces the bandwidth available for lower priorityflows. The second term is similar to the delay calcu-lated above, and reflects the effect of the new flow’sburstiness.

3. Effect of guaranteed flow on predictive traffic..Again, we compute the new delay bound � �-, for allpredictive classes directly from Eq. 1, adding in the

4

reserved rate, �45. , of flow . 87� � � �:�K��� ��L�=< > � � ���

������ < � � �JB>

�� ���=�� � ���

������ < ��L�=@� � �����:�K� < � � �?B> M �FR@QSRUT D(4)

Notice how the new guaranteed flow simplysqueezes the pipe, reducing the available bandwidthfor predictive flows; new guaranteed flows do notcontribute any delay due to their buckets becausethe WFQ scheduling algorithm smooths out theirbursts. Also observe that the first term of Eq. 3 isequivalent to Eq. 4: the impact of a new guaranteedflow is like adding a zero-size bucket, higher prior-ity, predictive flow.

Contrasting these three equations, we see that theexperienced delay of lower priority predictive traf-fic increases more when a higher priority predictiveflow is admitted than when a guaranteed flow or asame-priority predictive flow is admitted. The WFQscheduler isolates predictive flows from attempts byguaranteed flows to dump their buckets into the net-work as bursts. In contrast, lower priority predictivetraffic sees both the rates and buckets of higher pri-ority predictive flows. A higher priority predictiveflow not only squeezes the pipe available to lowerpriority traffic, but also preempts it.

B. Worst-case Delay: Guaranteed Service

In reference [Par92], the author proves that in anetwork with arbitrary topology, the WFQ schedul-ing discipline provides guaranteed delay boundsthat depend only on flows’ reserved rates and bucketdepths. Under WFQ, each guaranteed flow is iso-lated from the others. This isolation means that,as long as the total reserved rate of guaranteedflows is below the link bandwidth, new guaranteedflows cannot cause existing ones to miss their delaybounds. Hence, when accepting a new guaranteedflow, our admission control algorithm only needsto assure that (1) the new flow will not cause pre-dictive flows to miss their delay bound (see Eq. 4above), and that (2) it will not over-subscribe thelink: "4. 3 � 5. ��� � , where � is the link bandwidthand � is the utilization target (see Section III-B for adiscussion on utilization target). In addition to pro-tecting guaranteed flows from each other, WFQ also

isolates (protects) guaranteed flows from all predic-tive traffic.

C. Equivalent Token Bucket Filter

The equations above describe the aggregate traf-fic of each predictive class with a single tokenbucket filter. How do we determine a class’s to-ken bucket parameters? A completely conservativeapproach would be to make them the sum of theparameters of all the constituent flows; when datasources are bursty and flows declare conservativeparameters that cover their worst-case bursts, usingthe sum of declared parameters will result in lowlink utilization. Our algorithm is approximate andoptimistic: we take advantage of statistical multi-plexing by using measured values, instead of pro-viding for the worst possible case, to gain higherutilization, risking that some packets may occasion-ally miss their delay bounds. In essence, we de-scribe existing aggregate traffic of each predictiveclass with an equivalent token bucket filter with pa-rameters determined from traffic measurement.

The equations above can be equally described interms of current delays and usage rates as in bucketdepths and usage rates. Since it is easier to mea-sure delays than to measure bucket depths, we dothe former. Thus, the measured values for a predic-tive class � are the aggregate bandwidth utilizationof the class, �" � , and the experienced packet queue-ing delay for that class, �� � . For guaranteed service,we count the sum of all reserved rates, " . , and wemeasure the actual bandwidth utilization, �" . , of allguaranteed flows. Our approximation is based onsubstituting, in the above equations, the measuredrates �" � and �"%. for the reserved rates, and substi-tuting the measured delays �� �4$ ���*������� W for themaximal delays. We now use the previous compu-tations and these measured values to formulate anadmission control algorithm.

D. The Admission Control Algorithm

New Predictive Flow.. If an incoming flow re-quests service at predictive class ) , the admissioncontrol algorithm:

1. Denies the request if the sum of the flow’s re-quested rate, � 52 , and current usage would exceed

5

the targeted link utilization level:

� ��� �JB9 A � < > A�� �:��� � < � M (5)

2. Denies the request if admitting the new flowcould violate the delay bound, � 2 , of the same pri-ority level:

9 � � 9 A ��B9��� � @� � 9 ������� �< � M (6)

or could cause violation of lower priority classes’delay bound, � � : � � � � �L� � @� � ���

��:�K� � < ��L� � < > � � ��������� � < � � �?B9 A��B9��� � @� � ���

������ � < � � �JB9NM O�P@QSRUT D(7)New Guaranteed Flow.. If an incoming flow requests guaranteed service, the admission controlalgorithm:1. Denies the request if either the bandwidth checkin Eq. 5 fails or if the reserved bandwidth of allguaranteed flows exceeds the targeted link utiliza-tion level: � ��� �JB> A < > D (8)2. Denies the request if the delay bounds of pre-dictive classes can be violated when the bandwidthavailable for predictive service is decreased by thenew request:

� � � � ��� � @� � �����:�K� � < ��L� � < > � � ���

������ � < � � �JB> M �FR�QSRUT D(9)

If the request satisfies all of these inequalities, thenew flow is admitted.

III. A SIMPLE TIME-WINDOW MEASUREMENTMECHANISM

The formulae described in the previous sectionrely on the measured values �� � , �"%. , and �" � as in-puts. We describe in this section the time-windowmeasurement mechanism we use to measure thesequantities. While we believe our admission controlequations have some fundamental principles under-lying them, we make no such claim for the mea-surement process. Our measurement process usesthe constants � $�� , and ; discussion of their rolesas performance tuning knobs follows our descrip-tion of the measurement process.

A. Measurement Process

We take two measurements: experienced delayand utilization. To estimate delays, we measure thequeueing delay � of every packet. To estimate uti-lization, we sample the usage rate of guaranteed ser-vice, �"��. , and of each predictive class � , �"�� , overa sampling period of length � packet transmissionunits. Following we describe how these measure-ments are used to compute the estimated maximaldelay �� � and the estimated utilization �"%. and �" � .Measuring delay.. The measurement variable �� �tracks the estimated maximum queueing delay forclass � . We use a measurement window of packettransmission units as our basic measurement block.The value of �� � is updated on three occasions. Atthe end of the measurement block, we update �� �to reflect the maximal packet delay seen in the pre-vious block. Whenever an individual delay mea-surement exceeds this estimated maximum queue-ing delay, we know our estimate is wrong and im-mediately update �� � to be � times this sampled de-lay. The parameter � allows us to be more conser-vative by increasing �� � to a value higher than theactual sampled delay. Finally, we update �� � when-ever a new flow is admitted, to the value of pro-jected delay from our admission control equations.Algebraically, the updating of �� � is as follows:

���� �������� �������

������� ���� M of past T measurement window,� �� M if �� � � � ,Right sideof Eq. 6, 7,or 9,

when adding a new flow,depending on the serviceand class requested by theflow.

(10)

Measuring rate.. The measurement variables �" .and �" � track the highest sampled aggregate rate ofguaranteed flows and each predictive class respec-tively (heretofore, we will use “ �" ” as a shorthandfor “ �"%. and/or �" � ,” and “ �" � ” for “ �" �. and/or �" �� .”)The value of �" is updated on three occasions. Atthe end of the measurement block, we update �" toreflect the maximal sampled utilization seen in theprevious block. Whenever an individual utilizationmeasurement exceeds �" , we immediately update �"with the new sampled value. Finally, we update �"whenever a new flow is admitted. Algebraically, the

6

updating of �" is as follows:

� < � ����� ����������� �

not admit anymore flow until the end of a period.During its lifetime, � , a flow will see approximately� � � � � number of flows admitted every period.Thus at the end of its average lifetime, � , an averageflow would have seen approximately � � ��� � � number of flows. If the average rate of an averageflow is � � , ideally we want � � � � , a link’s stable uti-lization level, to be near � . However, flows alsodepart from the network. The expected number ofadmitted flow departures during the period de-pends on the number of flows and their duration. Ifthis number of departures is significant, a flow willsee a much smaller number of flows during its life-time, i.e. the stable � � � � becomes much smallerthan � . For the same average reservation rate, � ,and a given , the size of the stable � is determinedby the average flow duration, � . A shorter averageflow duration means more departure per . In thelong run, we aim for � � � ��� � , or equivalently,� � �� � � � � . If all flows use exactly what they re-served, we have � � � � , meaning that we shouldnot try to give away the flows’ reservations. Wepresent further illustrative simulation results on theimportance of the � � ratio in Section IV-E. Notethat when is infinite, we only use our computedvalues, which are conservative bounds, and ignorethe measurements entirely. That is, we will neversuffer any delay violations at a given hop if we usean infinite value for . Thus, the parameter al-ways provides us with a region of reliability.

IV. SIMULATIONS

Admission control algorithms for guaranteed ser-vice can be verified by formal proof. Measurement-based admission control algorithms can only be ver-ified through experiments on either real networks ora simulator. We have tested our algorithm throughsimulations on a wide variety of network topolo-gies and driven with various source models; we de-scribe a few of these simulations in this paper. Ineach case, we were able to achieve a reasonable de-gree of utilization (when compared to guaranteedservice) and a low delay bound violation rate (wetry to be very conservative here and always aim forno delay bound violation over the course of all oursimulations). Before we present the results fromour simulations, we first present the topologies andsource models used in these simulations.

HostA

Switch2Switch1

HostB

L1 L2L3

(a) One-Link

(b) Two-Link (c) Four-Link

HostA

Switch2Switch1

HostB

L1 L2L4

L5

Switch3L3

HostA

Switch1 Switch2

HostB

L1 L2

L7

L6

Switch3L3

Switch4

HostD

L9

L4

L8

HostE

Switch5

L5

HostC

HostC

Fig. 1. The ONE-LINK, TWO-LINK and FOUR-LINKtopologies

A. Simulated Topologies

For this paper, we ran our simulations on fourtopologies: the ONE-LINK, TWO-LINK, FOUR-LINK, and TBONE topologies depicted in Fig-ures 1(a), (b), (c), and 2 respectively. In thefirst three topologies, each host is connected to aswitch by an infinite bandwidth link. The con-nection between switches in these three topolo-gies are all 10 Mbps links, with infinite buffers.In the ONE-LINK topology, traffic flows fromHostA to HostB. In the TWO-LINK case, traf-fic flows between three host pairs (in source–destination order): HostA–HostB, HostB–HostC,HostA–HostC. Flows are assigned to one of thesethree host pairs with uniform probability. In theFOUR-LINK topologies, traffic flows between sixhost pairs: HostA–HostC, HostB–HostD, HostC–HostE, HostA–HostD, HostB–HostE, HostD–HostE;again, flows are distributed among the six host pairswith uniform probability. In Figure 1, these hostpairs and the paths their packets traverse are indi-cated by the directed curve lines.

The TBONE topology consists of 10, 45, and 100Mbps links as depicted in Figure 2(a). Traffic flowsbetween 45 host-pairs following four major “cur-rents” as shown in Figure 2(b): the numbers 1, 2, 3,4 next to each directed edge in the figure denote the“current” present on that edge. The 45 host-pairsare listed in Table I. Flows between these host-pairsride on only one current, for example flows fromhost H1 to H26 rides on current 4. In Figure 2(a),a checkered box on a switch indicates that we haveinstrumented the switch to study traffic flowing outof that switch onto the link adjacent to the check-ered box.

8

TABLE IFORTY-FIVE HOST PAIRS ON TBONE

Source Destination(s) Source Destination(s)

H1 H5, H7, H11, H14 H23 and H25H12, H14, and H26 H15 H11 and H17

H2 H10 and H25 H16 H5 and H9H3 H4 and H19 H17 H12H4 H18 H18 H5, H6, and H11H5 H14 and H25 H19 H5H6 H18 H20 H5H7 H17 H21 H9H8 H4, H5, H26 H22 H6H9 H3 and H19 H24 H12 and H17H10 H3 and H18 H25 H6 and H14H12 H4 H26 H9 and H14H13 H17 H27 H4

S13 S12 S8

S9

S10

S11

S4

S3S2

S5S6

S1S7H1

H2

H3

H4

H5H6H7

H8

H9

H10

H11

H12

H13

H14

H15

H16

H17

H18

H19

H20

H21

H22

H23

H24 H25

H26

H27

L1

L2

L3

L4

L5

L6

L7

L8

L9

L10

L11

L17

L18L19

L24

L26

L27

L36

L28

L29

L30

L31

L32

L33

L34

L35

L37 L38L39

L40

L12L13

L14L15

L16

L20L21

L22L23

L25

= Instrumentation

= 10 Mbps= 45 Mbps= 100 Mbps

(a) TBone topology

S13 S12 S8

S9

S10

S11

S4

S3S2

S5S6

S1S7H1

H2

H3 H4

H5H6H7

H8

H9

H10

H11

H12

H13

H14

H15

H16

H17

H18

H19

H20

H21

H22

H23

H24 H25

H26

H27

1

1 1

1

1 1

11

1

1

1

11

1

1

1

1

1

11

1

1

1

1

11

1

2 2

2

2

2

2

2

22

2

2

2

2

2

2

2

2

2

2

2

3

3

3

3

3

3

3

3 3

3

3

3 332

44

4

4

4

4

4

4

4

4

4

4

4

4

44

4

44

3

4

(b) Four traffic "currents" on TBone

Fig. 2. The TBONE topology

B. Source Models

We currently use three kinds of source model inour simulations. All of them are ON/OFF processes.They differ in the distribution of their ON time andcall holding time (CHT, which we will also call“flow duration” or “flow lifetime”). One of these isthe two-state Markov process used widely in the lit-erature. Recent studies ([LTWW94], [DMRW94],[PF94], [KM94], [GW94], [BSTW95]) have shownthat network traffic often exhibits long-range de-pendence (LRD), with the implications that con-gested periods can be quite long and a slight in-crease in number of active connections can result

in large increase in packet loss rate [PF94]. Refer-ence [PF94] further called attention to the possiblydamaging effect long-range dependent traffic mighthave on measurement-based admission control al-gorithms. To investigate this and other LRD relatedquestions, we augmented our simulation study withtwo LRD source models.

EXP Model.. Our first model is an ON/OFFmodel with exponentially distributed ON and OFFtimes. During each ON period, an exponentiallydistributed random number of packets, with aver-age

, are generated at fixed rate � packet/sec. Let�

milliseconds be the average of the exponentiallydistributed OFF times, then the average packet gen-eration rate � is given by � � � � ��� 3 � � � . TheEXP1 model described in the next section is a modelfor packetized voice encoded using ADPCM at 32Kbps.

LRD: Pareto-ON/OFF.. Our next model is anON/OFF process with Pareto distributed ON and OFFtimes (for ease of reference, we call this the Pareto-ON/OFF model). During each ON period, a Paretodistributed number of packets, with mean

and

Pareto shape parameter � , are generated at peakrate � packet/sec. The OFF times are also Paretodistributed with mean

�milliseconds and shape pa-

rameter � . Pareto shape parameter less than 1 givesdata with infinite mean; shape parameter less than2 results in data with infinite variance. The Paretolocation parameter is ������ � !�

������� ��� & �

������� .Each Pareto-ON/OFF source by itself does not gen-erate LRD series. However, the aggregation of themdoes [WTSW95].

LRD: Fractional ARIMA.. We use each num-ber generated by the fractional autoregressive inte-grated moving average (fARIMA) process ([HR89])as the number of fixed-size packets to be sent backto back in each ON period. Interarrivals of ON peri-ods are of fixed length. For practical programmingreasons, we generate a series of 15,000 fARIMAdata points at the beginning of each simulation.Each fARIMA source then picks an uniformly dis-tributed number between 1 and 15,000 to be usedas its index into that series. On reaching the end ofthe series, the source wraps around to the begin-ning. This method is similar to the one used bythe authors of [GW94] to simulate data from sev-

9

Tokensgeneratedat rate r

b tokens

Token BucketFilter

User Process

N packeton time

1/pM packetoff time

Host maxtransmissionrate C

C

Host

Network

Packets transmitted

at rate G, r

TABLE IISIX INSTANTIATIONS OF THE THREE SOURCE MODELS

Model’s Parameters Token Bucket Parameters Bound (ms)

Model Name � pkt/ � � ������

tkn/�

cut maxsec msec pkts sec tkns rate qlen

�EXP1 64 325 20 2 64 1 0 0 16 16EXP2 1024 90 10 10 320 50 2.1e-3 17 160 160EXP3 � 684 9 � 512 80 9.4e-5 1 160 160�POO1 64 2925 20 1.2 64 1 0 0 16 16POO2 256 360 10 1.9 240 60 4.5e-5 220 256 160

fARIMA( 0.75 � ,0.15, -)

� 125 8 13 1024 100 1.1e-2 34 100 160

time ( � ) of the Pareto-ON/OFF sources are selectedfollowing the observations in [WTSW95]. Accord-ing to the same reference, the shape parameter ofthe Pareto distributed OFF time ( � ) stays mostlybelow 1.5; in this paper we use � of 1.1 for allPOO sources. For the POO1 model, we use a to-ken bucket rate equals to the source’s peak rate suchthat the token bucket filter does not reshape the traf-fic. For the POO2 model, some of the generatedpackets were queued; this means during some of thesource’s alleged “OFF” times, it may actually still bedraining its data queue onto the network. Thus forthe POO2 model, the traffic seen on the wire maynot be Pareto-ON/OFF.

When a flow with token bucket parameters ! � $ � &requests guaranteed service, the maximal queue-ing delay (ignoring terms proportional to a singlepacket time) is given by

� �� [Par92]. Column 10

of the table, labeled � � , lists the guaranteed de-lay bound for each source given its assigned tokenbucket filter. Column 11, labeled � � , lists the pre-dictive delay bound assigned to each source. Wesimulate only two classes of predictive service. Apredictive bound of 16 msecs. means first class pre-dictive service, 160 msecs. second class. We havechosen the token bucket parameters so that, in mostcases, the delay bounds given to a flow by predic-tive and guaranteed services are the same. This fa-cilitates comparison between the utilization levelsachieved with predictive and guaranteed services.In the few cases where the delays are not the same,

such as in the POO2 and fARIMA cases, the utiliza-tion comparison is less meaningful. In the POO2case, for example, the predictive delay bound issmaller than the guaranteed bound, so the utiliza-tion gain we find here understates the true gain.

For the fARIMA source, we use an autoregres-sive process of order 1 (with weight 0.75) and de-gree of integration 0.15 (resulting in a generatedseries with Hurst parameter 0.65). The first orderautoregressive process with weight 0.75 means ourfARIMA traffic also has strong short-range depen-dence, while maintaining stationarity ([BJ76], p.53). The interarrival time between ON periods is1/8th of a second. The Gaussian innovation fed tothe fARIMA process has a mean of 8 packets withstandard deviation 13.

Except for simulations on the TBONE topology,flow interarrival times are exponentially distributedwith an average of 400 milliseconds. Because ofsystem memory limitation, we set the average flowinterarrivals of simulations on the TBONE topol-ogy to 5 seconds. The average holding time of allEXP sources is 300 seconds. The POO and fARIMAsources have lognormal distributed holding timeswith median 300 seconds and shape parameter 2.5.

We ran most of our simulations for 3000 secondssimulated time. The data presented are obtainedfrom the later half of each simulation. By visualinspection, we determined that 1500 simulated sec-onds is sufficient time for the simulation to warmup. However, simulations with long-range depen-

11

dent sources requesting predictive service requiresa longer warmup period. We ran all simulation in-volving such sources for 5.5 hours simulation time,with reported data taken from the later 10000 sec-onds.

We divide the remainder of this section up intothree subsections. First, we show that predictiveservice indeed yields higher level of link utilizationthan guaranteed service does. We provide support-ing evidence from results of simulations with bothhomogeneous and heterogeneous traffic sources, onboth single-hop and multi-hop networks. Depend-ing on traffic burstiness, the utilization gain rangesfrom twice to order of magnitude. This is the basicconclusion of this paper.

Second, we provide some simulation results to il-lustrate the effect of the � � ratio on network per-formance, as discussed in Section III-B. We showthat a larger � � ratio yields higher utilization butless reliable delay bound, while a smaller one pro-vides more stable delay estimate at lower utiliza-tion. We also present a few sample path snapshotsillustrating the effect of .

Finally, we close this section with a discussionof some general allocation properties of admissioncontrol algorithms when flows are not equivalent;we believe these properties to be inherent in all ad-mission control algorithms whose only admissioncriterion is to avoid service commitment violations.

D. On the Viability of Predictive Service

We considered six different source models, fourdifferent network topologies (one single hop andthree multi-hop), and several different traffic mixes.In particular, some of our traffic loads consisted ofidentical source models requesting the same ser-vice (the homogeneous case), and others had ei-ther different source models and/or different levelsof service (the heterogeneous case). The organiza-tion of our presentation in this section is: (1) ho-mogeneous sources, single hop, (2) homogeneoussources, multi-hop, (3) heterogeneous sources, sin-gle hop, and (4) heterogeneous sources, multi-hop.

Homogeneous Sources: The Single-hop Case..By homogeneous sources we mean sources that notonly employ just one kind of traffic model, but alsoask for only one kind of service. For this and

TABLE IIISINGLE-HOP HOMOGENEOUS SOURCES

SIMULATION RESULTS

Model Guaranteed Predictive

Name %Util #Actv %Util #Actv � � ��� � � �EXP1 46 144 80 250 3 60EXP2 28 28 76 75 42 300EXP3 2 18 62 466 33 600POO1 7 144 74 1637 5 60POO2 3 38 64 951 8 60fARIMA 55 9 81 13 72 60

all subsequent single-hop simulations, we use thetopology depicted in Figure 1(a). For each source,we ran two kinds of simulation. The first has allsources requesting guaranteed service. The secondhas all sources requesting predictive service. Theresults of the simulations are shown in Table III.The column labeled “%Util” contains the link uti-lization of the bottleneck link, L3. The “#Actv” col-umn contains a snapshot of the average number ofactive flows concurrently running on that bottlenecklink. The “ � ��� ” column contains the maximum ex-perienced delay of predictive class � packets. The“ � � ” column lists the ratio of average flow dura-tion to measurement window used with each sourcemodel.

We repeated the predictive service simulationsnine times, each time with a different random seed,to obtain confidence intervals. We found the con-fidence interval for the all the numbers to be verytight. For example, the utilization level of POO1sources under predictive service has a 99% con-fidence interval of (74.01, 74.19); the 99% confi-dence interval for the maximum experience delay is(4.41, 4.84) (the number reported in the table is theceiling of the observed maximum).

As mentioned in Section IV-B, we consider theperformance of our admission control algorithm“good” if there is no delay bound violation dur-ing a simulation run. Even with this very restric-tive requirement, one can see from Table III thatpredictive service consistently allows the networkto achieve higher level of utilization than guaran-teed service does. The utilization gain is not large

12

when sources are smooth. For instance, the sourcemodel EXP1 has a peak rate that is only twice itsaverage rate. Consequently, the data only shows anincrease in utilization from 46% to 80%. (One canargue that the theoretical upper bound in the uti-lization increase is the peak to average ratio.) Incontrast, bursty sources allow predictive service toachieve several orders of magnitude higher utiliza-tion compared to that achievable under guaranteedservice. Source model EXP3, for example, is a verybursty source; it has an infinite peak rate (i.e. sendsout packets back to back) and has a token bucketof size 80. The EXP3 flows request reservations of512 Kbps, corresponding to the token bucket rateat the sources. Under guaranteed service, only 18flows can be admitted to the 10 Mbps bottlenecklink (with 90% utilization target). The actual linkutilization is only 2%.4 Under predictive service,466 flows are served on the average, resulting in ac-tual link utilization of 62%.

In this homogeneous scenario with only one classof predictive service and constantly oversubscribedlink, our measurement-based admission control al-gorithm easily adapts to LRD traffic between thecoming and going of flows. The utilization in-creased from 7% to 74% and from 3% to 64%for the POO1 and POO2 sources respectively. Theutilization gain for the fARIMA sources was moremodest, from 55% to 81%. This is most proba-bly because the source’s maximum ON time is atmost twice its average (an artifact of the shiftingwe do, as discussed in Section IV-B, to obtain non-negative values from the fARIMA generated series).In all cases, we were able to achieve high levels ofutilization without incurring delay violations. Tofurther test the effect of long OFF times on ourmeasurement-based algorithm, we simulated POO1sources with infinite duration. With utilization tar-get of 90% link capacity, we did see a rather highpercentage of packets missing their delay bound.Lowering the utilization target to 70%, however,provided us enough room to accommodate trafficbursts. Thus for these scenarios, we see no reason toconclude that LRD traffic poses special challenges

�Parameter-based admission control algorithms may not

need to set a utilization target and thus can achieve a some-what higher utilization; for the scenario simulated here, twomore guaranteed flows could have been admitted.

TABLE IVMULTI-HOP HOMOGENEOUS SOURCES LINK

UTILIZATION

Link Model Guaranteed PredictiveTopology Name Name %Util %Util �������

EXP1 45 67 2L4 EXP3 2 44 20

POO2 3 59 7TWO-LINK

EXP1 46 78 3L5 EXP3 2 58 30

POO2 3 70 17

EXP2 17 42 6L6 POO1 4 31 1

fARIMA 38 54 36EXP2 28 71 31

L7 POO1 7 66 2fARIMA 55 77 40

FOUR-LINKEXP2 28 72 24

L8 POO1 8 75 7fARIMA 53 74 29EXP2 28 71 31

L9 POO1 8 59 2fARIMA 53 80 44

to our measurement-based approach.

Homogeneous Sources: The Multi-hop Case..Next we ran simulations on multi-hop topologiesdepicted in Figures 1(b) and (c). The top half ofTable IV shows results from simulations on theTWO-LINK topology. The utilization numbers arethose of the two links connecting the switches inthe topology. The source models employed here arethe EXP1, EXP3, and POO2 models, one per simu-lation. The bottom half of Table IV shows the re-sults from simulating source models EXP2, POO1,and fARIMA on the FOUR-LINK topology. For eachsource model, we again ran one simulation whereall sources request guaranteed service, and anotherone where all sources request one class of predictiveservice.

The most important result to note is that, onceagain, predictive service yielded reasonable lev-els of utilization without incurring any delay vio-lations. The utilization levels, and the utilizationgains compared to guaranteed service, are roughlycomparable to those achieved in the single hop case.

Heterogeneous Sources: The Single-hop Case..We now look at simulations with heterogeneoussources. For each of the simulation, we used twoof our six source model instantiations. Each sourcewas given the same token bucket as listed in Ta-ble II and, when requesting predictive service, re-quests the same delay bound as listed in the saidtable. We ran three kinds of simulation with hetero-

13

TABLE VSINGLE-HOP, SINGLE SOURCE MODEL, MULTIPLE

PREDICTIVE SERVICES LINK UTILIZATION

Model PP GP GPP

EXP1 77 77 –EXP2 71 70 –EXP3 31 31 –POO1 70 69 69POO2 60 57 –fARIMA 79 79 78

geneous sources: (1) single source model request-ing multiple levels of predictive service, (2) multi-ple source models requesting a single class of pre-dictive service, and (3) multiple source models re-questing multiple levels of predictive service. In allcases, we compared the achieved utilization withthose achieved under guaranteed service. For thefirst and third cases, we also experimented withsources that request both guaranteed and predic-tive services. When multiple source and/or servicemodels were involved, each model was given anequal probability of being assigned to the next newflow. In all these simulations, the experienced de-lays were all within their respective bounds.

Table V shows the utilization achieved whenflows with the same source model requested: twoclasses of predictive service (PP), guaranteed andone predictive class (GP), and guaranteed and twopredictive classes (GPP). In the GP case, flows re-quest the predictive class “assigned” to the sourcemodel under study (see Table II). In the othercases, both predictive classes, of bounds 16 and 160msecs. were requested. Compare the numbers ineach column of Table V with those in the “%Util”column of Table III under guaranteed service. Thepresence of predictive traffic invariably increasesnetwork utilization.

Next we look at the simulation results of multiplesource models requesting a single service model.Table VI shows the utilization achieved for selectedpairings of the models. The column headings namethe source model pairs. The first row shows the uti-lization achieved with guaranteed service, the sec-ond predictive service. We let the numbers speakfor themselves.

Finally in Table VII we show utilization num-

TABLE VISINGLE-HOP, MULTIPLE SOURCE MODELS, SINGLE

SERVICE LINK UTILIZATION

EXP1– EXP2– EXP2– EXP2– EXP3– POO2–Service POO1 EXP3 POO2 fARIMA fARIMA fARIMAGuaranteed 15 21 5 38 18 32Predictive 75 70 63 79 81 69

TABLE VIISINGLE-HOP, MULTIPLE SOURCE MODELS,

MULTIPLE PREDICTIVE SERVICES LINKUTILIZATION

EXP1– EXP1– EXP1– EXP2– EXP3– POO1–Service EXP2 fARIMA POO2 POO1 POO1 fARIMAGuaranteed 43 50 29 10 7 23Guar./Pred. 73 74 65 61 51 65Predictive 75 78 65 62 60 65

bers for flows with multiple source models request-ing multiple service models. The first row showsthe utilization achieved when all flows asked onlyfor guaranteed service. The second row shows theutilization when half of the flows requests guar-anteed service and the other half requests the pre-dictive service suitable for its characteristics (seeTable II). And the last row shows the utilizationachieved when each source requests a predictiveservice suitable for its characteristics.

Heterogeneous Sources: The Multi-hop Case..We next ran simulations with all six source mod-els on all our topologies. In Table VIII we showthe utilization level of the bottleneck links of thedifferent topologies. Again, contrast the utilizationachieved under guaranteed service alone with thoseunder both guaranteed and predictive services. Theobserved low predictive service utilization on linkL6 is not due to any constraint enforced by its ownadmission decisions, but rather is due to lack oftraffic flows caused by rejection of multi-hop flowsby later hops, as we will explain in Section IV-F.Utilization gains on the TBONE topology are notso pronounced as on the other topologies. Thisis partly because we are limited by our simula-tion resources and cannot drive the simulations with

14

TABLE VIIISINGLE- AND MULTI-HOP, ALL SOURCE MODELS,

ALL SERVICES LINK UTILIZATION

Topology Link Guaranteed Guaranteed and PredictiveName Name %Util %Util ������� ����� �ONE-LINK L3 24 66 3. 45.

L4 15 72 2. 54.TWO-LINK L5 21 72 2. 41.

L6 19 47 1. 36.L7 24 70 2. 46.

FOUR-LINK L8 20 72 2. 49.L9 18 75 1. 53.

L2 9 14 0.02 0.15L10 17 31 0.15 5.35L11 27 32 0.37 21.9

TBONE L12 22 23 0.1 5.84L20 8 21 0.22 16.6L30 32 52 0.49 34.7

higher offered load. Recall that flow interarrivals onsimulations using the TBONE topology have an av-erage of 5 seconds, which is an order of magnitudelarger than the 400 milliseconds used on the othertopologies.

Our results so far indicate that a measurement-based admission control algorithm can provide rea-sonable reliability at significant utilization gains.These conclusions appear to hold not just for singlehop topologies and smooth traffic sources, but alsofor multi-hop configurations and long-range depen-dent traffic as we have tested. We cannot, withinreasonable time, verify our approach in an exhaus-tive and comprehensive way, but our simulation re-sults are encouraging.

E. On the Appropriate Value of In Section III-B we showed that has two re-

lated effects on the admission control algorithm: (1)too small a results in more delay violations andlower link utilization, (2) too long a depresses uti-lization by keeping the artificially heightened mea-sured values for longer than necessary. While thefirst effect is linked to flow duration only if the flowexhibits long-range dependence, the second effectis closely linked to the average flow duration in gen-eral. The results in this section are meant to becanonical illustrations on the effect of on the ad-mission control algorithm, thus we do not providethe full details of the simulations from which theyare obtained.

In Table IX(a) we show the average link utiliza-tion and maximum experienced delay from simu-lations of flows with average duration of 300 sec-

TABLE IXEFFECT OF

�AND �

(a)�%Util � � ���

1e4 82 255e4 81 221e5 77 152e5 75 135e5 68 5

(b) �� 1e4 1e5

%Util � � � � %Util � � � �3000 86 48 82 24900 84 32 80 16300 82 25 77 15100 81 21 76 11

30 78 15 69 7

onds. We varied the measurement window, , from� � � packet times to � ��� packet times. Notice howsmaller yields higher utilization at higher expe-rienced delay and larger keeps more reliable de-lay bounds at the expense of utilization level. Nextwe fixed and varied the average flow duration.Table IX(b) shows the average link utilization andmaximum experienced delay for different values ofaverage flow duration with fixed at � � � and � ��� .We varied the average flow duration from 3000 sec-onds (practically infinite, given our simulation du-ration of the same length) to 30 seconds. Noticehow longer lasting flows allow higher achieved linkutilization while larger measurement periods yieldlower link utilization. Link utilization is at its high-est when the � � ratio is the largest and at its low-est when this ratio is the smallest. On the otherhand, the smaller � � ratio means lower experi-enced delay and larger � � means the opposite—thus lowering the � � ratio is one way to decreasedelay violation rate.

In Figures 4 and 5 we provide sample path snap-shots showing the effect of on delay and linkutilization. We note however, a that yields ar-tificially low utilization when used in conjunctionwith one source model may yield appropriate uti-lization when used with burstier sources or sourceswith longer burst time.

15

(a) Smaller (b) Larger

2035 2050

010

30

Simulated Time (secs.)

Del

ay (

mse

cs.)

Act

ual/M

easu

red

2035 2050

010

30

Simulated Time (secs.)

Del

ay (

mse

cs.)

Act

ual/M

easu

red

Fig. 4. Effect of�

on Experienced Delay

(a) Larger (b) Smaller

1880 1940 2000

01

2

Simulated Time (secs.)

Util

izat

ion

Mea

sdA

ctua

l/

# F

low

s74

90

1880 1940 2000

01

2

Simulated Time (secs.)

Util

izat

ion

Util

izat

ion

Mea

sdM

easd

Act

ual/

# F

low

s17

3Fig. 5. Effect of

�on Link Utilization

F. On Unequal Flow Rejection Rates

Almost all admission control algorithms in theliterature are based on the violation preventionparadigm: each switch decides to admit a flow ifand only if the switch can still meet all of its ser-vice commitments. In other words, the only criteriaconsidered by admission control algorithms basedon the violation prevention paradigm is whether anyservice commitments will be violated as a result of anew admission. In this section we discuss some pol-icy or allocation issues that arise when not all flowsare completely equivalent. When flows with dif-ferent characteristics—either different service re-quests, different holding times, or different pathlengths—compete for admission, admission con-trol algorithms based purely on violation preventioncan sometimes produce equilibria with some cate-gories of flows experiencing higher rejection ratethan other categories do. In particular, we iden-tify two causes of unequal rejection rate: (1) flowstraversing a larger number of hops have a higherchance of being rejected by the network, and (2)flows requesting more resources are more likely tobe rejected by the network.

Effect of Hop Count on Flow Rejection Rates. .As expected, when the network is as loaded asin our simulations, multi-hop flows face an in-creased chance of being denied service by the net-work. For example, in our simulation with homo-

geneous sources on the TWO-LINK network, as re-ported in Table IV, more than 75% of the 700 newEXP1 sources admitted under guaranteed serviceare single-hop flows. This is true for both of thebottleneck links. A somewhat smaller percentageof the more than 1000 flows admitted under pre-dictive service are single-hop flows. This effectis even more pronounced for sources that requestlarger amount of resources, e.g. the POO2 or thefARIMA sources. And it is exacerbated by sourceswith longer lifetimes: with fewer departures fromthe network, new flows see an even higher rejectionrate.

Aside from disparity in the kinds of flow presenton the link, this phenomenon also affects linkutilization; upstream switches (switches closer tosource hosts) could yield lower utilization thandownstream switches. We observe two causes tothis: (1) switches that carry only multi-hop flowscould be starved by admission rejections at down-stream switches. The utilization numbers of link L6in both Tables IV and VIII are consistently lowerthan the utilization of the other links in the FOUR-LINK topology. Notice that we set these simula-tions up with no single hop flow on link L6. Thelow utilization is thus not due to the constraint puton by link L6’s own admission decisions, but ratheris due to multi-hop flows being rejected by down-stream switches. (2) Non-consummated reserva-tions depress utilization at upstream switches; to il-lustrate: a flow admitted by an upstream switch islater rejected by a downstream switch; meanwhile,the upstream switch has increased its measurementestimates in anticipation of the new flow’s traffic,traffic that never come. It takes time (to the ex-piration of the current measurement window) forthe increased values to come back down. Duringthis time, the switch cannot give the reserved re-sources away to other flows. We can see this ef-fect by comparing the utilization at the two bottle-neck links of the TWO-LINK topology as reportedin Table IV. Note, however, even with the presenceof this phenomenon, the utilization achieved underpredictive service with our measurement-based ad-mission control algorithm still outperforms thoseachieved under guaranteed service.

16

Effect of Resource Requirements on Flow RejectionRates. . Sources that request smaller amount of re-sources can prevent those requesting larger amountof resources from entering the network. For exam-ple, in the simulation using the EXP2–EXP3 sourcepair reported in Table VI, 80% of the 577 new guar-anteed flows admitted after the simulation warmupperiod were EXP2 flows, which are less resource de-manding. In contrast, 40% of flows admitted underpredictive service with our measurement-based ad-mission control algorithm were the more resourcedemanding EXP3 flows. Another manifestation ofthis case is when there are sources with large bucketsizes trying to get into a high priority class. Be-cause the delay of a lower priority class is affectedby both the rate and bucket size of the higher prior-ity flow (as explained in Section II-A), the admis-sion control algorithm is more likely to reject flowswith a large bucket size and high priority than thosewith a smaller bucket size or low priority. We seethis phenomenon in the simulation of source modelEXP3 reported in Table V. When all sources re-quest either of the two classes of predictive servicewith equal probability, of the 1162 flows admittedafter the simulation warmup period, 83% were ofclass 2. When sources request guaranteed or sec-ond class predictive service, only 8% of the 1137new flows ends up being guaranteed flows. In bothof these scenarios, the link utilization achieved is31%, which is lower than the 62% achieved whenall flows request only class 2 predictive service (seeTable III), but still order of magnitude higher thanthe 2% achieved when all flows request only guar-anteed service (again, see Table III).

We consider the unequal rejection rate phe-nomenon a policy issue (or rather, several policyissues) because there is no delay violations andthe network is still meeting all its service commit-ments (which is the original purpose of admissioncontrol); the resulting allocation of bandwidth is,however, very uneven and might not meet somepolicy requirements of the network. We want tostress that this unequal rejection rate phenomenonarises in all admission control algorithms basedon the violation prevention paradigm. In fact, ourdata shows that these uneven allocations occur insharper contrast when all flows request guaranteed

service, when admission control is a simple band-width check. Clearly, when possible service com-mitment violations is the only admission controlcriteria, one cannot ensure that policy goals will bemet. Our purpose in showing these policy issuesis to highlight their existence. However, we do notoffer any mechanisms to implement various policychoices; that is the subject of future research andis quite orthogonal to our focus on measurement-based admission control.

V. MISCELLANEOUS PRACTICAL DEPLOYMENTCONSIDERATIONS

We have not yet addressed the issue of how toadjust the level of conservatism (through ) au-tomatically, and this will be crucial before suchmeasurement-based approaches can be widely de-ployed. The appropriate values of , and the otherparameters, must be determined from observed traf-fic over longer time scales than discussed (and sim-ulated) here. We have not yet produced such anhigher order control algorithm. In the simulationspresented in this paper, we chose a value of foreach simulation that yielded no delay bound viola-tion over the course of the simulation at “accept-able” level of utilization.

We should also note that our measurement-basedapproach is vulnerable to spontaneous correlationof sources, such as when all the TV channels aircoverage of a major event. If all flows suddenlyburst at the same time, delay violations will result.We are not aware of any way to prevent this kindof delay violation, since the network cannot predictsuch correlations beforehand. Instead, we rely onthe uncorrelated nature of statistically multiplexedflows to render this possibility a very unlikely event.

As we mentioned earlier, when there are onlya few flows present, or when a few large-grainflows dominate the link bandwidth, the unpre-dictability of individual flow’s behavior dictatesthat a measurement-based admission control algo-rithm must be very conservative. One may needto rely less on measurements and more on theworst-case parameters furnished by the source, andperform the following bandwidth check instead of

17

Eqn. 5:

� �����"%.=3������

�" �-$ (12)where,

�"%. � �"%.=3��! MAX !� $ "%.U� �"%. &H& $�" � � �" � 3� ! MAX !� $ " � � �" � &H& $ � � ������� W $"%. is the sum of all reserved guaranteed rates, " � isthe sum of all reserved rates in class � , W is numberof predictive classes, and is a fraction between 0and 1. For � � , we have the completely conser-vative case. Similarly, one could do the followingdelay check:

� � � �� ���� � 5�� � ��� � 5�� ��I"%.U���� ��������� " � � (13)

for every predictive class � for which one needs todo a delay check as determined in Section II-D.

VI. CONCLUSION

In this paper we presented a measurement basedadmission control algorithm that consists of twologically distinct pieces, the criteria and the esti-mator. The admission control criteria are based onan equivalent token bucket filter model, where eachpredictive class aggregate traffic is modeled as con-forming to a single token bucket filter. This enablesus to calculate worst case delays in a straightfor-ward manner. The estimator produces measuredvalues we use in the equations representing our ad-mission control criteria. We have shown that evenwith the most simple measurement estimator, it ispossible to provide a reliable delay bound for pre-dictive service. Thus we conclude that predictiveservice is a viable alternative to guaranteed servicefor those applications willing to tolerate occasionaldelay violations. For bursty sources, in particu-lar, predictive service provides fairly reliable delaybounds at network utilization significantly higherthan those achievable under guaranteed service.

APPRECIATIONS

This extended version of our ACM SIGCOMM’95paper [JDSZ95] has benefited from discussionswith Sally Floyd, Srinivasan Keshav, and Walter

Willinger; it has also been improved by incorporat-ing suggestions from the anonymous referees. Wethank them.

REFERENCES

[AM95] A. Adas and A. Mukherjee. “On Resource Man-agement and QoS Guarantees for Long RangeDependent Traffic. Proc. of IEEE INFOCOM1995, Apr. 1995.

[AS94] S. Abe and T. Soumiya. “A Traffic ControlMethod for Service Quality Assurance in an ATMNetwork”. IEEE Journal on Selected Areas inCommunications, 12(2):322–331, Feb. 1994.

[Ber94] J. Beran. Statistics for Long-Memory Processes.New York: Chapman & Hall, 1994.

[BJ76] G.E.P. Box and G.M. Jenkins. Time Series Analy-sis: Forecasting and Control. New Jersey: Pren-tice Hall, 1976.

[Bol94] V.A. Bolotin. “Modeling Call Holding TimeDistributions for CCS Network Design and Per-formance Analysis”. IEEE Journal on SelectedAreas in Communications, 12(3):433–438, Apr.1994.

[BSTW95] J. Beran, R. Sherman, M.S. Taqqu, and W. Will-inger. “Long-range Dependence in Variable-Bit-Rate Video Traffic”. IEEE Transactions on Com-munications, 43:1566–1579, 1995.

[C � 91] M. Conti et al. “Interconnection of Dual BusMANs: Architecture and Algorithms for Band-width Allocation”. Journal of Internetwork-ing: Research and Experience, 2(1):1–22, March1991.

[CLG95] S. Chong, S-Q. Li, and J. Ghosh. “Predictive Dy-namic Bandwidth Allocation for Efficient Trans-port of Real-Time VBR Video over ATM”. IEEEJournal on Selected Areas in Communications,13(1):12–23, Jan. 1995.

[Cru91] R.L. Cruz. “A Calculus for Network Delay, PartI: Network Elements in Isolation”. IEEE Trans-actions on Information Theory, 37(1):114–131,Jan. 1991.

[CSZ92] D.D. Clark, S.J. Shenker, and L. Zhang.“Supporting Real-Time Applications in anIntegrated Services Packet Network: Ar-chitecture and Mechanism”. Proc. of ACMSIGCOMM ’92, pages 14–26, Aug. 1992.URL ftp://parcftp.parc.xerox.com/pub/net-research/csz sigcomm92.ps.

[DJM97] Z. Dziong, M. Juda, and L.G. Mason. “A Frame-work for Bandwidth Management in ATM Net-works — Aggregate Equivalent Bandwidth Esti-mation Approach”. ACM/IEEE Transactions onNetworking, Feb. 1997.

[DKPS95] M. Degermark, T. Köhler, S. Pink, andO. Schelén. “Advance Reservations for PredictedService”. Proc. 5th Int’l Network and Operat-ing Systems Support for Digial Audio and VideoWorkshop, pages 3–14, Apr. 1995.

18

[DMRW94] D.E. Duffy, A.A. McIntosh, M. Rosenstein, andW. Willinger. “Statistical Analysis of CCSN/SS7Traffic Data from Working CCS Subnetworks”.IEEE Journal of Selected Areas in Communica-tion, 12(3):544–551, Apr. 1994.

[Flo96] S. Floyd. “Comments on Measurement-based Admissions Control for Controlled-Load Service”. Submitted to ComputerCommunication Review, 1996. URLftp://ftp.ee.lbl.gov/papers/admit.ps.Z.

[FV90] D. Ferrari and D.C. Verma. “A Scheme for Real-Time Channel Establishment in Wide-Area Net-works”. IEEE Journal on Selected Areas in Com-munications, 8(3):368–379, 1990.

[GAN91] R. Guérin, H. Ahmadi, and M. Naghshineh.“Equivalent Capacity and Its Application toBandwidth Allocation in High-Speed Networks”.IEEE Journal on Selected Areas in Communica-tions, 9(7):968–981, Sept. 1991.

[GKK95] R.J. Gibbens, F.P. Kelly, and P.B. Key. “ADecision-Theoretic Approach to Call AdmissionControl in ATM Networks.”. IEEE Journal onSelected Areas in Communications, 13(6):1101–1114, Aug. 1995.

[GW94] M. Garrett and W. Willinger. “Analysis, Mod-eling and Generation of Self-Similar VBR VideoTraffic”. Proc. of ACM SIGCOMM ’94, pages269–279, Sept. 1994.

[Hir91] A. Hiramatsu. “Integration of ATM Call Ad-mission Control and Link Capacity Control byDistributed Neural Network”. IEEE Journal onSelected Areas in Communications, 9(7):1131–1138, Sept. 1991.

[HLP93] J.M. Hyman, A.A. Lazar, and G. Pacifici. “ASeparation Principle Between Scheduling andAdmission Control for Broadband Switching”.IEEE Journal on Selected Areas in Communica-tions, 11(4):605–616, May 1993.

[Hos84] J.R.M. Hosking. “Modeling Persistence in Hy-drological Time Series Using Fractional Differ-encing. Water Resources Research, 20(12):1898–1908, Dec. 1984.

[HR89] J. Haslett and A.E. Raftery. “Space-time Mod-elling with Long-memory Dependence: Assess-ing Ireland’s Wind Power Resource”. AppliedStatistics, 38(1):1–50, 1989.

[Hui88] J.Y. Hui. “Resource Allocation for BroadbandNetworks”. IEEE Journal on Selected Areas inCommunications, 6(9):1598–1608, Dec. 1988.

[Jam95] S. Jamin. A Measurement-based Admis-sion Control Algorithm for Integrated ServicesPacket Network. Ph.D. Dissertation ProposalExcerpts. Technical Report USC-CS-95-617,Univ. of Southern California, CS Dept., 1995.URL http://netweb.usc.edu/jamin/admctl/quals-excerpts.ps.Z.

[JDSZ95] S. Jamin, P. B. Danzig, S. J. Shenker, andL. Zhang. “A Measurement-based AdmissionControl Algorithm for Integrated Services Packet

Networks”. Proc. of ACM SIGCOMM ’95, pages2–13, 1995.URL http://netweb.usc.edu/jamin/admctl/sigcomm95.ps.Z.

[JSZC92] S. Jamin, S.J. Shenker, L. Zhang, and D.D.Clark. “An Admission Control Algorithm forPredictive Real-Time Service (Extended Ab-stract)”. Proc. 3rd Int’l Network and OperatingSystems Support for Digial Audio and VideoWorkshop, Nov. 1992.URL http://netweb.usc.edu/jamin/admctl/nossdav92.ps.Z.

[Kel91] F.P. Kelly. “Effective Bandwidths at Multi-ClassQueues”. Queueing Systems, 9:5–16, 1991.

[KM94] S.M. Klivansky and A. Mukherjee. On long-range dependence in nsfnet traffic. Technical Re-port GIT-CC-94-61, Gerogia Insititue of Tech-nology, December 1994.

[LCH95] S-Q. Li, S. Chong, and C-L. Hwang. “LinkCapacity Allocation and Network Con-trol by Filtered Input Rate in High-SpeedNetworks”. ACM/IEEE Transactions onNetworking, 3(1):10–25, Feb. 1995. URLhttp://mocha.ece.utexas.edu/ sanqi/papers/link-cap.ps.

[LTWW94] W.E. Leland, M.S. Taqqu, W. Willinger, and D.V.Wilson. “On the Self-Similar Nature of EthernetTraffic (Extended Version)”. ACM/IEEE Trans-actions on Networking, 2(1):1–15, Feb. 1994.

[Mol27] E.C Molina. “Application of the Theory of Prob-ability to Telephone Trunking Problems”. TheBell System Technical Journal, 6:461–494, 1927.

[NK92] R. Nagarajan and J. Kurose. “On Defining, Com-puting, and Guaranteeing Quality-of-Service inHigh-Speed Networks”. Proc. of IEEE INFO-COM ’92, 1992.

[OON88] H. Ohnishi, T. Okada, and K. Noguchi. “FlowControl Schemes and Delay/Loss Tradeoff inATM Networks”. IEEE Journal on Selected Areasin Communications, 6(9):1609–1616, Dec. 1988.

[Par92] A.K. Parekh. A Generalized Processor SharingApproach to Flow Control in Integrated ServicesNetworks. PhD thesis, MIT, Lab. for Informa-tion and Decision Systems, Tech. Report LIDS-TR-2089 1992. Parts of this thesis were alsopublished with R.G. Gallager in the ACM/IEEETransactions on Networking, 1(3):344-357 and2(2):137-150.

[PF94] V. Paxson and S. Floyd. “Wide-Area Traffic: TheFailure of Poisson Modeling”. Proc. of ACMSIGCOMM ’94, pages 257–268, Aug, 1994. Anextended version of this paper is available asURL ftp://ftp.ee.lbl.gov/papers/poisson.ps.Z.

[SCZ93] S.J. Shenker, D.D. Clark, and L. Zhang.A Scheduling Service Model and aScheduling Architecture for an Inte-grated Services Packet Network. URLftp://parcftp.parc.xerox.com/pub/net-research/archfin.ps, 1993.

[SS91] H. Saito and K. Shiomoto. “Dynamic CallAdmission Control in ATM Networks”. IEEE

19

Journal on Selected Areas in Communications,9(7):982–989, Sept. 1991.

[WCKG94] R. Warfield, S. Chan, A. Konheim, and A. Guil-laume. “Real-Time Traffic Esitmation in ATMNetworks”. International Teletraffic Congress,June 1994.

[WTSW95] W. Willinger, M.S. Taqqu, R. Sherman, andD.V. Wilson. “Self-Similarity Through High-Variability: Statistical Analysis of Ethernet LANTraffic at the Source Level”. Proc. of ACM SIG-COMM ’95, pages 100–113, Aug. 1995.

[ZF94] H. Zhang and D. Ferrari. “Improving Utilizationfor Deterministic Service in Multimedia Com-munication”. IEEE Int. Conf. on MultimediaComputing and Systems, 1994.

[ZK94] H. Zhang and E.W. Knightly. “Providing End-to-End Statistical Performance Guarantee withBounding Interval Dependent Stochastic Mod-els”. Proc. of ACM SIGMETRICS’94, pages 211–220, 1994.

20