8/7/2019 Ubicc Paper 33 33

1/11

SCALABLE NETWORK BASED E2E ADMISSION CONTROL USING

FUZZY ARTMAP

R. Singh*, Verma S.**

*Dept. of C.S.I.T., I.E.T., MJP Rohilkhand University, Bareilly (India)(rsiet2002@gmail.com)** Indian Institute of Information Technology, Allahabad (India)

(sverma@iiita.ac.in)

ABSTRACTThe traditional approach to admission control assumes that traffic descriptor is provided by the user or application for each flow prior to establishment for the real timemultimedia communication. How ever this approach suffers from several problems. So,we propose a network-based endpoint admission Control system for scalable QoSguaranteed real-time communication services. This system is based on a sink tree-based

resource management strategy, and is particularly well suited for DiffServ based

architectures which keeps minimum signaling overhead by performing the admissiondecisions at the end points. In addition, the proposed system integrates routing andresource reservation along the routes, and therefore displays higher admission probabilityand better link resource utilization. This approach achieves low overall admission controloverhead because much of the delay computation is done during system configuration,and so resources can effectively be pre-allocated before run time. We investigate anumber of resources sharing approaches that allow resources to be efficiently re-allocatedat run time with minimized additional overhead. We provide simulation experiments thatillustrate the benefits of using DS tree-based resource management for resource pre-allocation and for routing, both with and without resource sharing.

Keywords: QoS, DiffServ, Scalability, DS-tree and RSVP.

1. INTRODUCTION

I N ORDER TO make quality-of-service (QoS)guarantees, a network must exercise flow admissioncontrol. Admission decisions are based on sometraffic characterization, such as effective bandwidth[7], [15] or leaky bucket descriptors [18]. Thereliable and efficient transmission of a real time

secure video on the Internet requires resources at the

server, client and network to be allocated in a

dynamic manner. It is generally accepted that end-to-end delay guarantees in networks and distributed

systems are provided by (i) appropriate allocation ofresources across the network, (ii) strict admission

control, and (iii) traffic monitoring and policing inthe network. This rather straightforward idea has been well studied, organized in the form of the Integrated Services architecture [3]. The mostserious shortcoming for the Integrated Servicesarchitecture is lack of scalability: All mechanismswithin the Integrated Services architecture rely on

flow awareness. The most natural way to cope with

this lack of scalability is to aggregate flows intoclasses of flows, and then have the network manageflow classes instead of individual flows. This general

approach is followed in the IETF DifferentiatedServices architecture [7, 12, 13]. The lack of per-flow information can negatively affect QoS

provisioning, as less information is available duringflow establishment and very limited policing can bedone during the flows lifetime. On the other hand, itallows for less expensive architectures for admission

control. As less flow information must be maintainedin the network, it is easier to centralize flowmanagement.

The admission control decision is made at a

central location for each administrative domain by bandwidth brokers [12], it significantly reduces the

cost of flow establishment, because only one entityin the domain needs to be contacted in the admission

control procedure. It is unlikely that this solution isscalable, however, given its tendency to have the

bandwidth broker node become a hot spot insituations with much flow establishment activity.

Performing admission control at the edge of thedomain can therefore further reduce theestablishment overhead. We call approaches thatlimit admission control activity to the edge of thedomain signaling-free admission control schemes.The signaling overhead of such approaches is either

literally zeros within the network, or it is sufficientlylight so as to not cause any scalability related

problem. Depending on where the admission control

is made, we classify signaling-free admission controlschemes into two categories: host-based andnetwork-based.

8/7/2019 Ubicc Paper 33 33

2/11

In a host-based signaling-free admission controlscheme, the host makes the admission decisionwithout invoking a signaling protocol. In a network-based signaling-free admission control the admissiondecision is performed by the ingress-router to thenetwork, and does not require signaling for the

decision when a flow arrives. So-called endpoint-based admission control mechanisms ([19, 20, 21,24]) typically fall under what we call the host-basedsignaling-free category.

Independently of where the decision is made, theadmission control has to have adequate and correctinformation about network resources at the time ofdecision. Otherwise, either delay guarantees are

violated, or the admission probability is unacceptablylow. Obviously, one way to eliminate the need for

signaling within the network is to control admission by probing the network at its edge. This is calledmeasurement-based admission control [24]. In this

approach, to set up a flow, the admission controlkeep sending the same rate of dummy traffic itwishes to set up for the predefined probing time.

After probing, the admission control makes thedecision whether it admits or not based on how thenetwork responded to the total traffic includingdummy packets. Literally, it requires zero signaling.However, no matter what technique is used in probing, the non-negligible side effect of long

latency for a flow set up makes it hard to be acceptedin real-time applications like Voice-over-IP. Thiswill be hardly acceptable because the users expectthe same or better quality of service when they have

experienced from traditional telephony systems.In this paper, we propose a completely different

signaling free admission control named network-based endpoint admission control, which: 1) requiresthe minimum possible signaling overhead, 2)

provides zero latency for a flow set up, 3) has zerorouting overhead, 4) has high admission probability,

and therefore high resource utilization, and 5) hastight end-to-end packet delay upper bounds. This isachieved by: 1) structuring available resources off-line using sink tree to reflect user trafficrequirements, 2) at run-time, limiting the admissioncontrol procedure to the ingress router only. The

ingress router then keeps track of the resourcesavailable downstream up to the destination. The restof the paper is organized as follows. Section 2describes previous work. In section 3, the DS treesystem is described. Section 4 presents resource-sharing strategies in the DS-tree paradigm. We

analyze the end-to-end delay in Section 5 for variousresource- sharing methods in the sink-tree paradigm

and simulated in section 6. Finally, conclusions andfuture work are described in Section 7.

2. PREVIOUS WORK

Some work has been done on admission control

of real-time applications with end-to-end delayconstraint within the DiffServ architecture [15, 22,23, 25]. The basic idea in [15] is to move per-flowinformation from core routers to edge routers byrelying on dynamic packet status carried with each packet. So the core router estimates each flows

dynamic information such as end-to-end delay foradmission control and packet scheduling based on

the dynamic packet status. The main idea of [25] isto apply the idea of the dynamic Packet State proposed in [15] to scalable admission control. Incontrast to these, work in [22] significantly reducesthe run-time overhead of admission control by doingsome of the computation off-line that is, during

network design or configuration. Other than these,the maximum achievable resource utilization in theDS model has been studied in [23], with a heuristicroute selection algorithm. While all of theseapproaches address the reduction of the overhead of

the admission control procedure proposed, they omitto address the overhead of the overall flowestablishment, of which admission control is just asmall part.

RSVP [2] has become the de facto of signalingfor resource reservation on the Internet, in particular

within the integrated services architecture [5].Recently, other resource reservation signaling protocols with reduced signaling overheads haveappeared, such as YESSIR [11], Boomerang [14],

and BGRP [17]. On the other hand, a series of effortshas focused on lightening the RSVP itself for

aggregated traffic [10, 16, 18]. These approaches still

cannot support a large number of real-timeapplications because 1) they still require rather longlatency in flow set-up because they rely on probing,and 2) they are not be able to guarantee bandwidthduring the service lifetime. For supporting of the

real-time applications in a scalable fashion, the softreservation paradigm is not appropriate.

In this paper we propose a network-basedendpoint admission control scheme that is scalable.The general idea of such a network-based admissioncontrol is presented in [26]. In that work we alsointroduce DS trees, discuss the problem of findingDS trees, the end-to-end delay analysis, and provide

basic simulation results on admission probabilities.In this paper, we address the practicality of the DStree-based approach. Specifically, we discuss how torelax rigid resource pre-allocation to respond tochanging flow establishment patterns while keepingsignaling overhead to a minimum.

3. THE DS TREE PARADIGM

For real-time applications resources must beallocated to the newly established flow and de-

allocated only after the flow has been torn down.Since flows require resources on a sequence of nodes

in the network, appropriate signaling must be in

8/7/2019 Ubicc Paper 33 33

3/11

place to synchronize the admission control.Independently of whether the signaling is centralized(such as in bandwidth-broker based approach) ordistributed (such as RSVP-style, for example), theoverhead in the case of high flow establishmentactivity is enormous. A scalable resource

management approach must therefore be able tomake admission decisions with high accuracy, whileavoiding both high message counts and centralizeddecision entities. Resource allocation overhead at runtime can be reduced by appropriately pre-allocatingresources during network re-configuration. When pre-allocating resources, four issues must beconsidered:

1) The allocation must reflect the expectedresource usage.

2) The allocated resources must be managed sothat the signaling overhead is minimized atrun time. Ideally, ingress routers manage pre-

allocated resources.3) Since the pre-allocation of resources defines

the routing of flows in the network, thesignaling must be appropriately integrated

with a routing mechanism.4) Pre-allocation commits resources early, and so

may result in low overall resource utilizationdue to fragmentation.

Lightweight mechanisms must be in place tochange the pre-allocation in order to accept flows

that could not be accepted in a system with rigidlycommitted resources. Allocating resources to DS

trees can satisfy the first three requirements.

Generally speaking (we give a more precisedefinition later,) since trees are used to aggregateconnections according to their egress nodes. The rootof a DS tree is then the egress router, and the leavesare the ingress routers. By allocating resources so

that each ingress router knows how much resource isahead for each path towards each egress router, the

admission control can be immediately performed atthe entrance of the network. Since each egress nodehas its own DS tree, every possible pair of sourceand destination node has its own unique path in a

sink tree. Consequently, wherever a flow arrives atan ingress node, the node determines the DS tree for

the new flow, based on the destination node. Thenthe path to the destination and the resources availableon the path is determined automatically. Anadmission decision can then be made at the very

node where a flow arrives. In the following weshortly discuss how typical network operations

would be performed with DS-trees.

N

= 1

B j l C l

Ingress Node Information: Each ingress node has

two types of information. One is the mapping table between the destination IP address in the input

packet and the corresponding IP address and portnumber of the egress router. The other is the tree

information corresponding to the egress router. Tree

information includes available bandwidth in the tree,the parent node in the tree, and the worst-casenetwork delay in the tree.

Admission Control at connection establishment,the connection initiator presents the admissioncontrollers with a connection request message, which

contains destination IP address, required bandwidth,and required network delay for the connection.

Secondly, there should be a connection tear downmessage. Finally, the input traffic assumed to beregulated by a leaky bucket at the traffic source.Once the ingress node receives a connection requestmessage, it looks up its routing table for thecorresponding egress router and port number. From

this information it determines which DS tree to use.Then it determines whether to admit or reject therequest based on available bandwidth and the worst-case network delay in that tree.

Packet forwarding once a connection is admitted,

a label is given to each input packet according to theDS tree it belongs to. So, each packet leaving theingress router for the parent has a new label in the packet. This label is used in packet forwarding in

core (internal) routers in the domain2 and is deletedwhen the packet leaves the egress router (root) for aneighboring domain. Consequently the label (treeID) is effective in a domain only. In other words, thecore routers do not see IP addresses; instead, theydeal with only the packet label for routing.

3.1. Model For Finding A Set of DS-Trees

Once we have a set of DS-trees for a givennetwork, we can run them as described in theprevious section. The problem of finding a sink treefor a given network could be formulated intheoretical graph design. We assume that the

network has no core routers. All the routers in thedomain are both ingress and egress routers. We also

assume that the network supports a single real-timeclass in addition to best-effort traffic. We define thedomain network as a graph G = {V,E} with nodes(routers or switches) in V, connected by links in E.Link l in E has link capacity Cl. For an output link jof an egress routers, we define a sink tree STj to be a

tree-like sub graph of G connecting Link j, whereeach link l in STj is marked with a bandwidthallocation Bl and k, which denotes the amount ofbandwidth allocated on Link l to real-time traffic onDS tree STj. For a set of DS trees {STj} to be valid,the following two constraints must be satisfied:

Link Capacity: On any link, the sum of bandwidthsallocated on the link for all DS trees should not

exceed the link capacity Cl. For each link, N

. (1)

8/7/2019 Ubicc Paper 33 33

4/11

Bjl Bjk

(K

.(2)

Where Bjl is the bandwidth allocated for the jth

DS

tree on link l, N is the total number of Ds-trees, and

Cl is the capacity of Link l.

Depth-aware Bandwidth Allocation: In a DS treeSTj, the bandwidth allocated on any link l should be

equal to or greater than the sum of bandwidthsallocated on all direct descendants of Link l in thesink tree.

(l, j))

Where (l, j) is the set of descendant links ofLink l in DS tree STj. In addition, the set of DS treesmust satisfy deadline requirements and bandwidth

requirements at the entrance and exit of the domain.These requirements are formulated in form of thefollowing three constraints:Bounded Network Delay:Every simple paths delay, which covers from an

ingress to an egress node, should be bounded by thegiven network delay.

MAX {Dp1,Dp2, . . . , Dpb} D (3)

Where Dpi is the worst-case delay in simple path pi, and D is the given network delay bound. SinceDpi = d1,2 + d2,3 + + dj-1,j , (for example, d1,2 is theworst-case packet forwarding delay from node 1 to

node 2) Dpi is the sum of worst-case local delays ineach simple path.

Bandwidth at Domain Entrance: For each ingressrouter, the amount of requested bandwidth for eachpath from that router to each egress router should beallocated as requested.

ri,j = Bi,j .. (4)

Where, ri,j is the requested bandwidth for the pathbetween Router i, and Router j. BBi,j is the allocatedbandwidth for the path by the DS tree.

Bandwidth at Domain Exit: For each egress router,the sum of bandwidth allocated for the direct

descendant links of the DS tree (the egress router)should not exceed the given output bandwidth forthat DS tree.

Where, Tj is the given output bandwidth for DS

tree j. (j) is the set of descendant links of the egressrouter. We showed in [26] that the problem of

finding a valid set of DS trees for a given network is

NP-Complete in all but the most trivial cases. Wealso described a heuristic algorithm based on MST

(Minimum Spanning Tree), which performs verywell in finding valid DS trees if such DS trees exist.If it fails to find a valid DS tree, the heuristicalgorithm returns a set of reduced bandwidthrequirements at the network entrance and relaxeddeadline requirement so that a valid set of DS trees

can be found. Unless otherwise mentioned, we usethis heuristic algorithm (Fuzzy ARTMAP) for thefollowing simulations study.

4. RESOURCE SHARING IN THE DS TREE

A fixed partitioning of resources and theirallocation to DS trees gives rise to fragmentation,which can significantly reduce overall resourceutilization, when the flow population along source

paths exceeds the expected values. Re-computationof resource allocations is very expensive in terms ofcomputation time and communication overhead to

distribute the information about the newconfiguration for ingress routers. A lightweightmethod is needed to allow for adaptive re-allocation

of resources between reconfigurations. In this sectionwe describe how to take advantage of the resourceallocation structure in sink tree to share resourceswithin portions of the DS tree or across multiple DStrees.

The general idea behind resource sharing isillustrated in Figure 1. As seen in the figure, the four

nodes (Nodes 1, 2, 3, and 4) are on the simple path ina DS tree. Node 4 is the DS (root), and other

branches are not drawn. Each link should be

allocated resources so that each node canaccommodate the equal number of real time flowsfrom it through the sink. So, Node 1 should be ableto accept ten real-time flows from Node 1 to Node 4.Likewise, Node 2 should be able to accept ten real-

time flows from Node 2 to Node 4, and so on.Accordingly, the links have bandwidth allocations

ranging from 10 to 30 from left to right so that Link1 supports ten real-time flows, Link 2 supportstwenty, and so on. Now suppose that Node 1 has onereal-time flow to Node 4, while Node 2 has ten real-

time flows to Node 4. In this situation, Node 2receives the 11th real-time flow set-up request. At

this point, resources can be shared in three differentways.

Tj Bjk

(K (j))

... (5)

Figure 1: An illustration of resource sharing

8/7/2019 Ubicc Paper 33 33

5/11

)(

)()1()(

+++=

N

YTYd

No Sharing: In this strategy resources are not shared.Bandwidth on each link is exclusively allocated to a pair of source and destination. In the example inFigure 1, the 11th admission request at Node 2 isdenied even though the path from Node 2 to Node 4(Link 2 and Link 3) has resources available for 9

new requests in total. We will use no sharing asbaseline for comparison with other sharing strategies.

kkk

...(6)

=

pathj

jSpath

k dYk

max ..(7)

( =u ulB1 , )

Path Sharing: In path-sharing resources are sharedalong overlapping paths of the same DS tree. In ourexample, the 11th admission request is admittedbecause the path from Node 2 to Node 4 (Link 2 andLink 3) has still resources available. As a side effect,

if Node 2 has admitted 20 flows, it starves Node 1.So the resource is shared only along overlapping

paths of the same tree.

Tree Sharing: Assume that in our example there is

no available bandwidth in the tree to accept the 11thflow at Node 2. With tree-sharing the admissioncontrol attempts to use bandwidth allocated to othersink trees on the path from Node 2 to Node 4. As

long as there is such a tree, the 11th request isaccepted. In this approach, resource is sharedbetween trees that share the same set of consecutivelinks.

Link Sharing: In this approach, the new request is

admitted as long as there are resources available onany tree along the requested path regardless of which

tree the resource originally belongs to. The DS tree

only determines the path for a flow and the resourcesare shared among all tree edges in the domain.

5. END-TO-END DELAY ANALYSIS

In this section we investigate the DS-trees effecton the worst-case end-to-end delay in the presence of

resource sharing.

No-Sharing Case: In order to calculate the worst-case end-to-end network delay, we model thenetwork as a collection of servers [4]. Once we getan upper bound on the delay on each server, we can

easily obtain the end-to-end delay by simplysumming up all the worst-case local delays. Forsimplicity of analysis, we consider a special case inwhich we have only two classes of traffic: real-timewith priority and non-real-time best-effort trafficsuch that the real-time traffic is never delayed by nonreal-time best-effort traffic. In addition, all flows inthe real-time class have the same traffic

characterization in terms of bandwidth and leakybuckets parameters.

Finally the input real-time traffic to the network

is constrained by (, ) [1]. Inside the networkrouters are not flow-aware, therefore there is no

traffic regulator.

Theorem 1:

Where

In our case, a formula for the local delay dk atServer k can be formulated as follows based on [1, 4,6, 22]. is the burst size of source traffic of the flowin the real time class. Sk is the set of all possiblepaths upstream from Server k for flows that traverseServer k. Yk is the maximum delay a flow

experienced before arriving at Server k. N is thenumber of input links to Server k. is the portion of

the link resource allocated for the real-time class

traffic. Because the upper bound has been derivedunder the assumption that the worst-case local delaycomes with the maximum workload, there is no

dynamic, run-time dependent variable in Equation(6). If we assume that Yk is a constant (of course this

is not true, i.e., each node has different values of Yk),the upper bound is a monotone increasing function of. This is exactly correspondent to the commonsense that the higher the workload, the larger theend-to-end delay.

Sharing Case: In an attempt to compare the worst-case end-to-end delays for the different resource

sharing strategies, we first represent themathematically according to the strategies, becausethe bandwidth for the real time class is limited by on each link. Followings are the notations ofbandwidths for the representation of.

=

+

=

+++++=

t

uii

ull

tlululllul

BC

BBBBCC

,1

,

,1,1,1,, )(

Where, Cl represents the capacity of Link l. Since

we consider only the two classes in this paper, anylink capacity consists of two parts: one for real-timepriority class

Best-effort basis ( BBbest-effort ). In other words, Link lhas t number of tree-edges, each of them has its ownallocated resource and the rest of the resource ofLink l is for the best-effort traffic. Cl,u means the

resource available for the edge of DS -tree u on Link

and the other for the traffic by

n

ul

d

ululul RRRB ,,1

,, ++++=

..(8)

..(9)

8/7/2019 Ubicc Paper 33 33

6/11

l. Since the real-time class has priority, C l,u has twotypes allocations: one for the resource for the best-effort traffic on Link l and the other for the resourceallocated for DS-tree u in Link l. Bl,u is the resourceallocated to DS-tree u on Link l. R

dl,u is resource

requirement for the path from Node d to the sink of

DS-tree u which passes Link l. Now four variants of are given below using these notations.

l

t

u ul

t

l

n

u

n

u

uul

t

ul

d

ul

p

ul

d

ul

n

C

B

CC

BB

C

B

C

R

=

= =

=

++

++

=

=

=

1 ,

1

1 1

,,

,

,

,

,

The above four equations (n, p, t, l) are for nosharing, path sharing, tree sharing, and link-sharing

respectively. These equations are ratherstraightforward according to the definitions of

sharing strategies in the previous section. In t,

means the number of trees that share a set of linksunder consideration. Based on the common sense, asmaller value of will get a tighter value of theupper bound. Moreover, since we know the paths theflows will take, we can take advantage of it in

calculating end-to-end delay. However, the fourequations of reveal little about the order ofmagnitude among them. Because the values of andYk are determined in the DS tree constructionprocess, and we know the final values only when thetrees are constructed, they cannot be comparedbefore the construction is completed. In comparisonof Equation (11) and (12), the inequality is obvious

because p is always not smaller than n, becauseR

dl,u is expected to less than or equal to B l,u at best.

The worst case end-to-end delay obtained here isused in the heuristic algorithm for constructing DS-trees proposed in the next section.

6. ALGORITHM DESIGN ISSUES FOR

CONSTRUCTING DS TREES

Although finding a set of DS-trees for a givennetwork under the three constraints is NP-Complete

[26], we need a practical algorithm which can always produce a set of DS-trees in the polynomial time

which may or may not have the reduced

requirements in both bandwidth and network delay.So, we tried to achieve by a Fuzzy ARTMAPalgorithm is, to always produce a set of sink trees inthe polynomial time with the minimum networkdelay and with or without reduced constraints for agiven network. With the link capacity, we think that

it is usually large enough to accommodate the pathbandwidth delays. However, because we do not tryto find an optimal path value, which might beminimized reduction in requirement, we do not sayeither our heuristic algorithm produces best possibleDS-trees or it guarantees performance with a certain percentage of difference from the optimal. The

reduction has two cases: One is for link capacity andother is for network delay.

(11)

.(12)

6.1. RESOURCE ALLOCATION PROBLEM

The video stream needs adequate resources at the

server, network and the client with constraints inDiffServ architecture. It is processed at the source processor, transmitted over the network and then processed at the destination. However, the cost of

the corresponding resource results in lower or higherutility for the DS trees. The algorithm always stopswith a DS tree for each out port. So if the givennetwork has t number of output ports, the algorithmiterates exactly t times. Since we wish to constructDS-tree in the polynomial time, we limit the

execution of the loops for just one time for eachviolation such that the loop execution are bounded by number of edges |E| for each DS-tree. This

apportioning of the resources has to be in real timeand the allocation time is a part of the serializationdelay to develop DS-trees. The ARTMAP [27] iselastic enough able to either find an existing cluster

or plastic enough to form a new QoS clustercorresponding to a SLA.

.(13)

(14)

6.2 ARTMAP ARCHITECTURE

The ARTMAP neural network [28] is anarchitecture that can learn arbitrary mappings fromdigital inputs of any dimensionality to outputs of anydimensionality. The architecture applies incremental

supervised learning of recognition categories andmultidimensional maps in response to arbitrarysequences of analog or binary input vectors, whichmay represent or crisp set of features. The ARTMAPneural network consists of two ART modulesdesignated as ARTa and ARTb, as well as an interART module. Inputs are presented at the ARTamodule, while the outputs are presented at the ARTbmodule. The Inter ART module includes a MAP

field, whose purpose is to determine whether themapping between the presented input and output is

the desired one. The vigilance parameter constraints,the maximum size of the compressed representations

of the input/output patterns. The range of the

8/7/2019 Ubicc Paper 33 33

7/11

vigilance parameter is the interval [0,1]. Small valuesof the vigilance parameter, results in coarseclustering of the patterns, while large values lead tofine clustering. ART is a clustering algorithm thatoperates on vectors with discrete-valued elements.Adding a further layer of processing (the MAP

field) to ART yields a supervised clusteringalgorithm. During learning, ARTMAP is presentedwith training vectors that have been labeledaccording to the class to which each belongs. TheART module assigns the current training vector to acluster.

If this procedure, termed match tracking,

successfully assigns the training vector to a cluster

with the correct class association, the weightsdefining that cluster are updated, making it morelikely to capture that training vector in the future. Ifno cluster with the correct class association is found,a new cluster is created.

7. EXPERIMENTAL EVALUATION

In this section, we evaluate the performance ofthe proposed network-based endpoint admissioncontrol by simulation experiments that can beguaranteed by each strategy to evaluate the DS-treeseffect on the end-to-end delay. The performance is

measured in terms of worst-case end-to-end delay,maximum possible resource allocation, admission probability, and resource utilization. Finally, weinvestigate how this performance can be affected bythe configuration of a given network. We compareour approachs performance with that of a strategythat uniformly allocates a fixed amount of bandwidthto every link of the given network (We call this

strategy flat-fixed). This strategy shall use ShortestPath First (SPF) routing in setting-up a flow. Theperformance comparison will eventually be betweenthe two systems: One allocates resources evenly toeach link and selects paths using SPF when a flow

request arrives, and the other uses the DS-tree

approach for both resource allocation and pathselection. The network-based endpoint admissioncontrol proposed in this paper cannot be compared

directly to any host-based endpoint admissioncontrol in terms of performance, because the latter

mainly focuses on the probing technique, which

cannot guarantee both end-to-end delay andbandwidth for the flows lifetime. First of all, Figure2 shows the topology of the MCI ISP-backbone

network, which we use throughout the experiments.In the first set of experiments, all routers (black and

gray) can act as edge routers and core routers as well.This means that the flows can arrive and leave atevery router. In a second set of experiments, blacknodes are edge routers and gray nodes are corerouters. We consider two classes of traffic in thissimulation study:

ARTMAP Algorithm:

Step1:Application, User & Network parametersare given;

Step2: Out port = t, Iterate for each output port;

Step3: Construct the DS-tree, for each constraint;

Step4:If DS-tree not constructed thenStep5:If link capacity violated then

Reduce the bandwidth requirement,

And Go to step 3;

Step6: else Network delay violated thenReduce the Network delay requirement

And Go to step 3;

End;

Step7: else output is the DS-tree;

End;

Figure 2: A network example

One real-time class and one non-real time class.We assume that all flows in the real-time class havea fixed packet length of 640 bits (RTP, UDP, IPheaders and 2 voice frames) [8], and a flow rate of

32 Kbps. The end-to-end delay requirement of allflows is fixed at 100ms. Thus, the input traffic ofeach flow is constrained by a leaky bucket withparameters = 640 bits and = 32 kbps. This kindof QoS requirements is similar to Voice-over IP.All links in the simulated network have the samecapacity of 155 Mbps.

For the DS-tree system we assume that the

bandwidth requested for any path between a pair ofingress and egress nodes remains the same. In other

words, for a nodes point of view, it has the sameamount of bandwidth available to every egress node.In this experiment, the amount of resource for a path between any two nodes is 1.28 Mbps, which

accommodates 40 real-time flows. For the flat-fixedsystem, roughly 10.5% of the link capacity is

allocated for the real-time flows. So, every link canaccommodate 510 real-time flows. The numbercomes from the fact following. After the constructionof DS-trees, the total sum of resources allocated to

each link in the DS-tree system divided by the

number of total links gives roughly 16.3 Mbps,

8/7/2019 Ubicc Paper 33 33

8/11

=t

u

which corresponds to 10.5% of 155 Mbps-capacitated link.

Item Flat-Fixed

No-

Sh.

Path

-Sh.

Tree

-Sh.

Link-

Sh.

Worst case

e2e delay 3.138 1.609 4.204 4.204 13.580

Table 1: Worst-case end-to-end delay for eachresource-sharing strategy (msec.)

Worst-case End-to-end Delay: Table 1 shows theworst-case end-to-end delay (unit : msec.) in eachresource sharing strategy. As can be seen, the no-sharing strategy outperforms flat-fixed, while linksharing is significantly worse. We interpret this asfollows. According to the equations about

(Equations (11) to (14)), the end-to-end delaybecomes larger with increasing bandwidth allocated

over the link for real-time traffic, while it getssmaller when the link capacity grows. The reason forthe smallest value for no-sharing is that Rdl,u is muchsmaller than B

dl,u. In the case of link-sharing

d

ulB1 ,

to produce the worst (largest) value. In flat-fixedsystems the resources are distributed evenly over the

network. Flat-fixed therefore, produces smaller end-to-end delays than link-sharing. So in terms of performance, sink-tree paradigm, except for link-sharing, provides similar or tighter worst-case end-

to-end delays.

Maximum Possible Resource Allocation: We definethe maximum possible resource allocation as theratio of total resource allocated for the real-timetraffic over the total capacity of all links under thecondition that the worst-case end-to-end delayreaches up to the worst case end-to-end delay

requirement (Table 2). In the table, UBAC stands forUtilization Based Admission Control which waspresented in [23].

Item

Flat-

Fixed

UBAC no-

sh.

Path

Sh.

Tree

-Sh.

Link

-Sh.

Alloc 0.33 1.609 0.57 0.49 0.49 0.39

Table 2: Maximum possible resource allocation foreach resource-sharing strategy.

This approach illustrates the benefit of good pathselection: For each possible pair of source anddestination nodes, the paths are predefined off-line,and among them, the most promising one (that

provides the smallest end-to-end delay) is selected asruntime path. The bandwidth is allocated uniformly

over the all links, however. UBAC therefore

provides the limit on maximum possible bandwidthallocation for the flat-fixed system with optimalrouting. It is important to note here that all proposednetwork-based endpoint admission control systems,except for the link-sharing strategy, outperform the

UBAC system. As can be seen, UBAC gives about

50% performance improvement from the flat-fixedsystem. However, no-sharing provides 70%improvement from flat-fixed. As expected, as weshare more resources, the maximum possibleresource allocation gets smaller since the sharingcauses loose end-to-end delay.

Admission Probability: We simulate the admissioncontrol behavior in the system by simulating flowrequests and establishments at varying rates with a

constant average flow lifetime. Requests for flowestablishment form a Poisson process with rate ,while flow lifetimes are exponentially distributed

with an average lifetime of 180 seconds for eachflow. Source and destination edge routers are chosenrandomly. In the flat-fixed system, the path is

selected using SPF routing. In the DS-tree systemhowever, the path is pre-defined in a DS-tree. Figure3 shows the admission probabilities for the real-timeclass in the five cases as a function of arrival rates. Inthe figure, the legends are that CA1-flat for CallAdmission-probability with Configuration 1 (black

and gray nodes are all edge and core nodes) in theflat-fixed system, CA1-no for Call Admission- probability with Configuration 1 in the no-sharingsystem and so on.

-may be large enough on some links

Figure 3: Admission probabilities

Call-admission-probability

0 20 40 60 80 100 120 140 160

call-arrival- rate

1

.9

.8

.7

.6

.5

.4

As can be seen in Figure 3, the four sink-tree

systems perform better than flat-fixed. Even no-sharing outperforms flat-fixed. In the case of no-

sharing, if we wish to allow 90% chances of flowadmission, the performance is approximately 50%

larger than that of flat-fixed. The importance of thisresult is that the sink-tree structured resourcemanagement brings a lot of benefit even when all thebandwidth requirements, and all link capacity are the

same. The differences between the four resource-

8/7/2019 Ubicc Paper 33 33

9/11

sharing strategies in the sink-tree system are notnegligible either. At 90% chances of flow admissionpoint, link-shared strategy outperforms no-sharing byabout 10%, with a little more signaling overhead. Ifthis signaling overhead is light enough to scale, thisimprovement is important.

Figure 4: Link utilization

Figure 4 shows the utilization ratios for the five

cases. To be fair, the link utilization is defined bythe resource consumed by the flows in the networkdivided by the total resources allocated. So this linkutilization is different measure of performance from

the maximum possible resource allocation describedearlier in this section. The former effectively

measures the efficiency of allocation, while the

latter provides a limit on the maximum allowableresources under the worst-case end-to-end delayconstraint. By comparing Figure 4 and Figure 3, weknow that the point of 90% admission probability inno-sharing system corresponds to 75-calls-per-

second. At this call arrival rate, the flat-fixed systemutilizes resources only around 70%, while no-sharingdoes around 85%. In other words, even no-sharingstrategys efficiency is better than that of the flat-fixed system by about 20% in link resourceutilization. The point is that although the total

amount of allocated resources are the same, theutilization depends on how the resources are

allocated. Therefore the admission probability will be better with an efficient resource allocation. The big difference between 50% improvement inadmission probability and 20% improvement in

resource utilization comes from the fact that thenumber of links requested by a call ranges from 1 to

the longest length of the simple path in the DS-tree.This accommodates more calls requesting the smallfragment of resources rather than the calls requestinglonger paths (many number of links). So the numberof calls accepted gets higher. In overall, the proposednetwork-based endpoint admission control shows

significant improvement in admission probability

over the flat-fixed system. One thing, which might

look confusing in Figure 3, is that the admissionprobability of link-sharing goes down below that of

tree-sharing at around 80% chances of admissionpoint. We answer this with Figure 5. As can be seen,the number of links per accepted call (averageresources per accepted call, which is acronym-ed

NOL in the figure) is decreasing in all resource-sharing strategies. Interestingly, the link-sharings

NOL goes over that of tree-sharing. This must haveresulted in the higher admission probabilities fortree-sharing by the same argument (resourcefragmentation) mentioned above.

In fact, we do not put an emphasis on this

because in most real cases, we expect the networkwill be administered at 90% or more chances ofadmission points. We consider this as a network

configuration-specific result.

Network-configuration Effect on Resource Sharing:To illustrate the effect, we ran the same set ofsimulations with Configuration 2, where the core

nodes are not the edge nodes any more. That means

Number-of-links-

per-accepted-call

0 20 40 60 80 100 120 140 160

call-arrival- rate

Figure 5: Average amount of resource peraccepted call

Call-admis

sion-probability

0 20 40 60 80 100 120 140 160

call-arrival- rate

1

.9

.8

.7

.6

.5

.

4

Figure 6: Admission probabilities with

Configuration 2

Resource-utilization-rate

1

.

9

.8

.7

.6

.5

.4

.3

.

2

.1

0 20 40 60 80 100 120 140 160

call-arrival- rate

0

8/7/2019 Ubicc Paper 33 33

10/11

only the edge nodes can be the sources ordestinations for the flow. For this particularexperiment, the black nodes in Figure 2 are the edgenodes (1,2,3,4,5,6,7,8,9,0), while the gray nodes arecore nodes (10,11,12,13,14,15,16,17,18). Followingthree figures 6, 7, and 8 show the admission

probabilities, resource utilization rates, and averageresources per accepted call respectively.

Figure 7: Link utilization with Configuration 2

Figure 8: Average amount of resource peraccepted call with Configuration 2

By comparing Figure 3 and Figure 6, we observe

that with Configuration 2, the admission probabilitiesare higher and the differences between the resource-sharing strategies are smaller. For example, the linksharing has 10% improvement from no-sharing inFigure 3, while it has only 3% improvement inFigure 6. The reason is that because only the edgerouters can be sources and destinations the resourcesare less fragmented than in Configuration 1. In otherwords, in Figure 1, Node 2 and 3 are not edge nodes

anymore, so there is no flow request at these nodeswith Configuration 2. Therefore the benefit ofsharing is reduced by the same argument, there isonly a little difference in link resource utilization in

Figure 7 too.

In overall, it is clear that there is only a littledifference in the admission probabilities andresource utilization among the four resource-sharingstrategies with Configuration 2. We expect that if wehave more pure core routers, the difference will bemuch smaller. Observing these results, we can say

that if the number of edge nodes is relatively smallercompared to the number of total nodes, the resource

sharing does not contribute much for admissionprobability or resource utilization.

0 20 40 60 80 100 120 140 160

call-arrival- rate

resource-utilization-rate

1

.9

.8

.7

.6

.5

.4

.3

.2

.1

0

8. CONCLUSIONS

In this paper, we proposed a network-basedendpoint admission control which: 1) requires theminimum possible signaling overhead, 2) provides

minimum possible latency for a flow set up, zerorouting overhead, high admission probability, highresource utilization, and tight end-to-end packet

delay upper bound. This is achieved by : 1) havingresources structured off-line with DS tree reflectingthe user traffic requirement, 2) at run-time, referring

only to the edge router (the entrance of the network)at which a flow arrives and which automaticallykeeps track of the resources available downstream upto the destination. This approach is more suitable tothe many real-time applications like Voice-over-IPthan the host-based endpoint control. For evaluation

purposes we compared this with the flat-fixed systemwhere all links have the same capacity and a fixedportion of the link capacity is allocated for the real-time traffic over the network uniformly with SPF

routing.Simulation study shows that the proposed

system with Fuzzy ARTMAP algorithm provides50% improvement in the tightness of the worst-caseend-to-end delay. Also, in terms of maximum possible resource allocation, it provides 75%improvement, both from the flat-fixed system

respectively. Simulation studies on the admissionprobabilities show that no resource-sharing strategyoutperforms the flat-fixed system by up to 50%improvement. Also, we showed the effect of networkconfiguration on the resource-sharing by simulationswith different network configurations. According to

the result, there is not much benefit of resourcesharing if the number of edge nodes is smallercompared to the number of total nodes in the givennetwork. In the future, we will extend the DS-tree paradigm on a global scale so that the very largenetwork will be able to support real-time applications

in a scalable fashion.

REFERENCES:

[1] R. L. Cruz, A Calculus for Network Delay,

Part I and Part II, IEEE Trans. on InformationTheory, Jan. 1991.

0 20 40 60 80 100 120 140 160

call-arrival- rate

2.4

2.35

2.3

2.25

2.2

2.15

2.1

2.05

2

1.95

Number-of-links-per-accepted-call

8/7/2019 Ubicc Paper 33 33

11/11

[2] L. Zhang, S. Deering, D. Estrin, S. Shenker andD. Zappala, RSVP: a new resource reservation protocol, IEEE Networks Magazine, vol. 31,No. 9, pp. 8-18, September 1993.

[3] R. Braden, D. Clark and S. Shenker, IntegratedServices in the Internet Architecture, RFC

1633, Jun. 1994.[4] A. Raha, S. Kamat, W. Zhao, Guaranteeing

End-to-End Deadlines in ATM Networks, The15th IEEE International Conference onDistributed Computing Systems, in 1995.

[5] P. P. White, RSVP and Integrated Services inthe Internet: A Tutorial, IEEECommunications Mag., May 1997.

[6] C. Li, R. Bettati, W. Zhao, Static PriorityScheduling for ATM Networks, IEEE Real-

Time Systems Symposium (RTSS97), SanFrancisco, CA. Dec. 1997.

[7] K. Nicols, V. Jacobson, L. Zhang, A Two-bit

Differentiated Services Architecture for theInternet, Internet-Draft, Nov. 1997.

[8] Thomas J. Kostas, Michael S. Borella, IkhlaqSidhu, Guido M. Schuster, Jacek Grabiec, and

Jerry Mahler, Real-Time Voice Over Packet-Switched Networks, IEEE Network Mag,Jan./Feb. 1998.

[9] T. Ferrari,W. Almesberger, J. L. Boudec, SRP:a scalable resource reservation protocol for theInternet, In Proceedings of IWQoS, pp. 107-

117, Napa CA. May. 1998.[10] A. Terzis, L. Zhang, E. Hahne, Reservations

for Aggregate Traffic: Experiences from an

RSVP Tunnels Implementation, The 6th IEEEInternational Workshop on Quality of ServiceNapa Valley, CA, May 1998, pp. 23-25.

[11] P. Pan, H. Schulzrinne, YESSIR: A SimpleReservation Mechanism for the Internet, In theProceedings of NOSSDAV Cambridge, UK,Jul. 1998.

[12] S. Blake, D. Black, M. Carlson, E. Davies, Z.Wang, W. Weiss, An Architecture forDifferentiated Service, RFC 2475, Dec. 1998.

[13] Y. Bernet et al., A Framework forDifferentiated Services, Internet-Draft, IETF,Feb. 1999.

[14] G. Feher, K. Nemeth, M. Maliosz,Boomerang: A Simple Protocol for ResourceReservation in IP Networks, IEEE Real-TimeTechnology and Applications Symposium, June1999.

[15] I. Stoica, H. Zhang, Providing GuaranteedServices without Per Flow Management. ACMSIGCOMM, Sep. 1999.

[16] L.Wang, A. Terzis, L. Zhang, A New Proposal

for RSVP Refreshes, The 7th IEEEInternational Conference on Network Protocols

(ICNP99), Oct. 1999.[17] P. Pan, E. L. Hahne, H. Schulzrinne, BGRP: A

Tree-Based Aggregation Protocol for Inter-

Domain Reservations, Bell Lab TechnicalMemorandum, Work Project No.MA30034004, MA30034002, File Case 20564,Dec. 1999.

[18] Y. Bernet, The Complementary Roles of RSVPand Differentiated Services in Full-Service QoS

Network, IEEE Communications Mag., Feb.2000.

[19] G. Bianchi, A. Capone, C. Petrioli,Throughput analysis of end-to-endmeasurement-based admission control in IP, InProceedings of IEEE INFOCOM, Tel Aviv,Israel, Mar. 2000.

[20] C. Cetinkaya, E. Knightly, Egress admissioncontrol, In Proceedings of IEEE INFOCOM,Tel Aviv, Israel, Mar. 2000.

[21] V. Elek, G. Karlsson, R. Ronngren, Admissioncontrol based on end-to-end measurements, InProceedings of IEEE INFOCOM, Tel Aviv,

Israel, Mar. 2000.[22] B. Choi, D. Xuan, C. Li, R. Bettati, W. Zhao,

Scalable QoS Guaranteed CommunicationServices for Real-Time Applications, The 20th

IEEE International Conference on DistributedComputing Systems, Taipei, Taiwan, pp. 180-187, April 2000.

[23] D. Xuan, C. Li, R. Bettati, J. Chen, W. Zhao,Utilization-Based Admission Control for Real-Time Applications, The IEEE International

Conference on Parallel Processing, Canada,Aug. 2000.

[24] L. Breslau, E. W. Knightly, S. Shenker, I.

Stoica, H. Zhang, Endpoint AdmissionControl: Architectural Issues and Performance,In Proceedings of ACM SIGCOMM 2000Stockholm, Sweden, Aug.-Sep. 2000.

[25] Z. Zhang, Z. Duan, L. Gao, Y. T. Hou,Decoupling QoS Control from Core Routers:A Novel Bandwidth Broker Architecture for

Scalable Support of Guaranteed Services, InProceedings of ACM SIGCOMM 2000Stockholm, Sweden, Aug.-Sep. 2000.

[26] B. Choi, R. Bettati, Efficient ResourceManagement for Hard Real-TimeCommunication over Differentiated Services

Architectures, The 7th IEEE InternationalConference on Real-Time Computing Systemsand Applications, Cheju, Korea, Dec. 2000.

[27] R. Singh, Ph. D. Thesis on Study and Analysisof Some QoS issues in the Internet, LucknowUniversity, Lucknow (India).

[28] A Carpenter, S. Grossberg, N. Markuzon, H.Reynolds, and D. B Rosen, "Fuzzy ARTMAP:A neural network architecture for incremental

supervised learning of analog multidimensionalmaps," IEEE Transactions on Neural Networks,

vol. 3, no. 5, Sept. 1992, pp. 698-712.