Wireless Networks 0 (1998) ?{?
1
Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance Anup Kumar Talukdar a; B. R. Badrinath a ; Arup Acharya b a
Department of Computer Science, Rutgers University, New Brunswick, NJ 08903 E-mail: ftalukdar@paul,
[email protected] b C&C Research Labs, NEC USA, Princeton, NJ 08540 E-mail:
[email protected]
This paper considers the support of real-time services to mobile users in an Integrated Services Packet Network. In the currently existing architectures, the service guarantees provided to the mobile hosts are mobility dependent, i.e. mobile hosts experience wide variation in the Quality of Service and often service disruption when hosts move from one location to another. The network performance degrades signi cantly when mobile hosts are provided with mobility independent service guarantees. In this paper we have proposed a service model for mobile hosts that can support adaptive applications which can withstand service degradation and disruption, as well as applications which require mobility independent service guarantees. We describe an admission control scheme for implementing this service model and evaluate its performance by simulation experiments. Simulation results show that, if sucient degree of multiplexing of the mobility dependent and independent services are allowed, the network does not suer any signi cant performance degradation and in particular our admission control scheme achieves high utilization of network resources. Keywords: Integrated Services Packet Networks, mobility, Quality of Service, admission control, performance
1. Introduction Recent progress in computing technology and wireless digital communication has made portable computers easily available. This has led to an intensive research in the area of Mobile Computing to provide mobile users access to an inter network. The research, so far, has focussed on the problem of maintaining connectivity at the network and transport layer in spite of the mobility of the mobile hosts[12,4,3]. Also, there have been several proposals for supporting real-time applications in an Integrated Services Packet Network (ISPN). Typical applications that require real-time services include audio library, image browsing, video conferencing and video-on-demand. These multimedia applications require a bound, which may be absolute or statistical, on the delivery delay of each packet. Clark et. al.[5] have described an architecture for an ISPN that supports real-time trac. In
This research work was supported in part by DARPA under contract numbers DAAH04-95-1-0596 and DAAG55-97-1-0322, NSF grant numbers CCR 95-09620, IRIS 95-09816 and Sponsors of WINLAB.
this architecture they have described a service model for real-time trac with two service classes, guaranteed and predictive. Jamin et. al.[7] have described a scheme to implement the service model described in [5]. As portable computers become more powerful and the accessibility of a xed network from a mobile host becomes easier, the number of mobile users will grow and additionally the mobile users will demand the same real-time services which are available to xed hosts. Some of the applications which require real-time services in a mobile environment are Internet Cellular Phone[2], and Call Center Applications in which oline information retrieval over the internet can be integrated with on-line interactive services. In this paper, we investigate the problems of providing real-time services to mobile hosts in an ISPN. As described in [5], to handle real-time services an enhanced architecture is required for a network. This architecture has ve key components. The rst component is the nature of service commitments the network can provide. The second component is the service interface parameters passed between the network and the ow endpoints; this includes both the characterization of the quality of
2
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
service the network can provide and the characterization of the behavior of the endpoints of the ow. The third component is the functionality of the network elements, namely scheduling algorithms, required to meet the service commitments. The fourth component is the admission control mechanism by which the service commitments are established at each network element. The last component is the reservation protocol to setup the required states in the network elements for providing the required service. In this paper, we discuss how mobile users can be accommodated within this architecture1 . This paper is organized as follows. In Section 2 we discuss the problems of providing real-time services to mobile users in an ISPN. In Section 3 we outline the related work in this area and describe our approach to the problem. In Section 4 we describe the service interfaces and propose the service model required for the ISPN with mobile users and describe the scheduling algorithms used for the service classes; Section 5 describes the admission control mechanisms for our proposed service model and outline the mobile hando techniques; in Section 6 we brie y outline a resource reservation protocol for our architecture; Section 7 contains discussion on the performance results; we conclude in Section 8 with some future directions of this work.
2. Real-time services to mobile users: Problems In an ISPN, the main Quality of Service (QoS) parameter is the delay experienced by a packet traveling from a sender to a receiver. The packet delay has three components. The rst component, propagation delay, is the propagation time of the packet at the speed of light; this is xed once the data path is chosen. The data path, usually, consists of multiple hops, connected by switches or routers. The second component is the delay in transmission at each switch waiting for the entire packet to arrive before it can be transmitted onto the next link; this delay depends on the packet size. The third component, congestion delay, arises due to the statistical sharing of the nite link bandwidth by packets belonging to multiple ows: packets arriving at a switch are buered in a service queue until the outgoing link is available. Congestion delay of a packet is the 1
A preliminary version of this paper appeared in The Proceedings of The INFOCOM 1997, Kobe, Japan, April 1997[15].
time spent by it in the service queue; this delay depends on the number of ows using a link, the link capacity and the trac generation rate of dierent sources. To provide real-time services, this congestion delay must be bounded or minimized. Clark et. al.[5], have classi ed the real-time applications into two classes, tolerant applications which can adapt to packet delays and thus can tolerate occasional delay bound violations, and intolerant applications which cannot tolerate any delay bound violations. Correspondingly, they have de ned two service classes: guaranteed service for intolerant applications and predictive service for tolerant applications. In an ISPN consisting of static hosts QoS guarantee is provided by reserving sucient bandwidth at each link along the path from the sender to the receiver for the duration of the session, so that the congestion delay at each switch is bounded. In a mobile computing network, a geographical area is divided into several cells. Each cell is served by a base station (also called mobility agent) connected to the xed network. A mobile host maintains connectivity with the xed network through the base station of the cell in which it is currently located. When a mobile host moves from one location to another, the delivery delay of a packet is aected in two ways. The rst factor is, the propagation delay may change due to the change in the length of the path from the sender to the receiver. Secondly, the congestion delay at the switches along the new path may be dierent. Therefore, when a mobile host initiates a session with a certain QoS guarantee by reserving link bandwidth along the path from the sender to its current location, the QoS guarantee is valid only in that location; when the mobile host moves to a new location, the QoS guarantee is not valid in the new location. There are two broad ways to provide service guarantees to mobile hosts. The rst approach is location dependent: the QoS is guaranteed speci c to a location, i.e., QoS guarantee is maintained as long as the mobile host stays at the location from where it initiated the session. To obtain such a service the mobile host makes a reservation for a certain QoS from its current location. As soon as it moves to a new location, the QoS guarantee is no longer valid. The application has to renegotiate its desired QoS at the new location. If sucient resources are not available along the new data
ow path, the mobile host may suer service degra-
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
dation, which we call hando failure. The second approach is location independent: the mobile host receives the same QoS guarantee at all locations (given by a mobility pro le2 ) where it may move during the lifetime of the session. For example, suppose at time t1 , the mobile host is at cell C1 where it initiates a session with a certain QoS guarantee, and at time t2 it is located in cell C2 where t1 < t2 . It will receive the same QoS guarantee at time t2 as long as C2 is within its mobility pro le. To obtain mobility independent QoS, a mobile host must make reservations along the data ow path from the sender to each location it may visit. However this leads to very low utilization of resources because, although data is physically owing to all locations, data is being used only at the current location of the mobile. The utilization of the network resources can be signi cantly improved if the reserved resources of unused data
ows could be used by other ows. However, the users who are utilizing these unused resources may suer service degradation when the original reservers start using these resources. For example, suppose a mobile host M1 has made reservation at the two cells, C1 and C2 , and it is currently located at C1 and hence using the reserved resources at C1 . In this situation, another host M2 which is located in C2 can use the resource reserved by M1 in C2 . But when M1 moves into C2 , M2 must release the resources reserved for M1 at C2 .
3
tion of the link bandwidth to enhance hando success for mobile hosts. Levine et. al.[9] have proposed resource allocation and admission control schemes based on a new concept, called Shadow Cluster, to improve the QoS of mobile calls by reducing the number of dropped calls in an wireless-ATM network. Lu et. al.[10] have described an adaptive resource management algorithm for indoor mobile environments. These research have focussed on providing QoS guarantees to adaptive applications which can withstand wide range of available bandwidth. In these works, the main objective of the admission control and resource allocation schemes is to reduce the rate of hando failures of the mobile hosts without signi cant degradation in the utilization level of the network bandwidth. Mobile applications subscribing to those services may suffer variable degrees of QoS degradation as they move from one location to another. Therefore these schemes are not suitable for more rigid applications which require QoS guarantees that are not aected by mobility of hosts. In this paper we propose an architecture that accommodates applications which are adaptive to variable degradation in QoS and also those applications which require mobility independent QoS guarantees in the same network. A salient feature of this architecture is that, a mobile host is required to negotiate its desired QoS only at the start of a data ow session and eliminates the need of fresh reservation every time it moves 3. Related work and our approach to a new location during the lifetime of the session. The admission control scheme, we propose, attempts to inRecently there has been some work addressing the crease utilization of network bandwidth by statistical problem of providing QoS to mobile hosts. Acampora multiplexing of link bandwidth among ows of dierent and Naghshineh[1] propose an architecture for a highclasses of applications. speed mobile ATM network using a new concept known as the virtual connection tree and describe an analytical method of admission control for mobile hosts. Singh[14] 4. Service Interfaces and Service Model has introduced two new QoS parameters, loss pro le and probability of seamless communication, arising due To obtain a certain QoS in an ISPN, the endpoints to the mobility of the mobile hosts and described a net- of a data ow is required to provide characterization of work architecture and a suite of transport level services the data trac they will generate, so that the network to satisfy these QoS parameters. Lee[8] has presented can reserve sucient resources for the ow. This charan \adaptive reserved service" framework for use in in- acterization is done with a token bucket lter[5,17]. A tegrated services networks to support mobile connec- token bucket lter has two parameters: its token genertions carrying multimedia trac. They have proposed ation rate, r, and the depth of its bucket, b. Tokens are an adaptive link partitioning scheme to reserve a frac- generated at the xed rate r and accumulate into the 2 mobility pro le of a mobile host is the set of locations the mobile bucket. When the bucket becomes full, i.e. contains b host is expected to visit tokens, newly generated tokens are discarded. When a
4
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
packet of size p is generated, it can be released into the network only after removing p tokens from the bucket. In addition to securing the required number of tokens from the bucket the trac is constrained to enter the network at a maximum rate of C > r, called peak rate. A trac ow conforms to a token bucket lter (r; b) if there are enough tokens in the bucket whenever a packet is generated. In a xed network the required resources are reserved on the links along the data ow path. In a mobile environment the data ow path changes as a mobile host moves from one cell to another. To provide real-time services to a mobile host, it is necessary to reserve resources along all possible data paths which may be used during the lifetime of a connection. Therefore, in addition to the trac characterization, the mobile host is also required to provide its mobility characterization to the network. In this paper we assume that, the mobility of an user is predictable so that mobility can be characterized precisely by mobility speci cation which consists of the set of cells the mobile host is expected to visit during the lifetime of the ow. In the rest of this section we describe the service model for the mobile multimedia applications and the scheduling algorithms used for this service model. 4.1. Service Model
The service model, we propose, is derived from the service model, containing the two service classes guaranteed and predictive services, described in [5] by Clark et. al. Guaranteed service provides a hard or absolute bound on the delay of every packet and is suitable for those applications which cannot tolerate any violation of delay bound. However, many recently developed applications for packet switched networks, such as vat, nv and vic, can adapt to actual packet delays and are thus tolerant of occasional delay bound violations. These tolerant applications can be served by predictive service, which oers a fairly, but not absolutely, reliable bound on packet delays. We de ne three service classes to which mobile users may subscribe. These are:
its trac characterization. This class is appropriate for the intolerant applications which require absolute bounds on packet delays in spite of mobility of the hosts. 2. Mobility Independent Predictive service (MIP class): A mobile user admitted to this class will receive predictive service with respect to packet delay bound as long as its moves are limited to its mobility speci cation and it is conforming to its trac characterization. This class is appropriate for those tolerant applications which require fairly reliable delay bounds in all cells it might visit but does not want to be severely aected by mobility of the hosts. 3. Mobility Dependent Predictive service (MDP class): A mobile user admitted to this service class will receive predictive service in all cells it may visit during the lifetime of its connection as long as its moves are limited to its mobility speci cation and it is conforming to its trac characterization. However it may occasionally fail to get predictive service and experience severe degradation of QoS and may be dropped when they move to a new location or some other ow of class MIG or MIP moves into a link on its data ow path. This class is appropriate for tolerant applications, which can tolerate the eects of service degradation due to mobility of hosts.
The predictive ows (both in MIP and MDP classes) are categorized into a number of levels with widely separated delay bounds. The delay bounds of the corresponding levels of MIP and MDP ows are same, i.e. MIP level j and MDP level j have the same delay bounds. To implement the above service model for mobile hosts in ISPN, it is not enough to reserve resources along the path from the sender to the current location of the mobile host; it is necessary to make reservations along the paths to other locations where the mobile host may visit. However, it is not necessary to initiate the data ow along each of those paths; data ow is initiated only along the path from the sender to the current location of the mobile host. Thus we de ne two types 1. Mobility Independent Guaranteed service (MIG of ows: class): A mobile user admitted to this service class will receive guaranteed service with respect to 1. Active : A ow is active on a link along a data path packet delay bounds as long as its moves are limited if resources are reserved for the ow on the links to its mobility speci cation and it is conforming to along the data path and data is being transmitted
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
5
to a receiver along that path. The corresponding teed service is provided by the Weighted Fair Queuing reservation is called an active reservation. scheduling discipline described in [11,6] and the predictive service is provided by the priority queuing algo2. Passive : A ow is passive on a link along a data path if resources are reserved for the ow on the rithm using the bandwidth leftover from serving guaranteed service. The ows within each predictive level links along the data path but the data is not passing (both MIP and MDP service classes) are scheduled by through the link. The corresponding reservation is FIFO algorithm. called a passive reservation. One approach to reservation in an ISPN is to use the a priori worst-case trac characterization of each reserving ows to compute the resource requirements to provide the requested services. However, if the ow trac is bursty, the average ow rate is signi cantly less than their a priori worst-case characterization and as a result the network utilization will be very low. To achieve higher utilization, Jamin et. al.[7] have proposed a measurement-based admission control scheme for predictive services. This scheme uses a priori characterization of new ows, but uses measurement to characterize existing ow. They have shown signi cant gain in utilization when there is a high degree of multiplexing. In our proposed reservation scheme, we use the above measurement-based admission control algorithm. For a
ow of MIG or MIP class, the required amount of bandwidth is reserved along the paths from the sender to all locations in the mobility speci cation of the receiving mobile host. However the ow is active only along the path to the current location. To increase the utilization of the network bandwidth, we allow the ows of the MDP class to use the reserved but unused resources of the passive reservation of MIG and MIP classes. When a passive ow becomes active on a link (i.e. the mobile who made the passive reservation moves into the cell), some MDP ow may suer QoS degradation if the MDP ows were using the unused resources of the passive ow. 4.2. Scheduling Algorithm
From a scheduling point of view, the ows on a link are divided into two classes: guaranteed and predictive. The MIG ows belong to the guaranteed service class; the MIP and MDP ows belong to the predictive class. There are multiple levels of predictive service with widely separated delay bounds. The scheduling algorithm considers only the active ows on a link. The uni ed scheduling algorithm of [5] is used, i.e., guaran-
5. Admission Control Scheme The goal of the admission control algorithm should be to admit as many mobile users as possible with the requested QoS and achieve a very high utilization of the resources. However, when a new mobile user is admitted, the system should ensure that it meets its prior commitments to previously admitted users. To admit a new ow it is necessary to perform admission control at all switches and base stations along the paths from the sender to the cells where the mobile host may visit. The admission control criteria at a switch for a ow will depend on the following two factors: service class and whether the ow is active or passive on the outgoing link of the switch when the ow starts. A ow is admitted to the network only if it passes the admission control criteria at each switch or base station where the
ow requires reservation. Our admission control algorithm is derived from the measurement-based admission control scheme for ISPN proposed by Jamin et. al.[7]. In Reference [7], Jamin et. al. have described a technique to measure the queuing delays, link usage rates of the guaranteed service class and each predictive class. These quantities have been used to compute, for each predictive class, the equivalent token bucket lter that characterizes the currently existing aggregate trac pattern of the active
ows of that class. In their scheme, the maximal delay bounds of predictive ows are approximated by replacing the worst-case parameters in the analytical model with equivalent token bucket lter. A new ow is admitted only if it can get the requested service and the delay bounds of the existing ows are not violated after admitting this new ow. The admission control criteria in our service model will use the same equivalent token bucket lter parameters whenever possible. In addition, we augment the measurement process of [7] to measure the token generation rate of the equivalent token bucket lter for each individual active ow. Before describing
6
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
the admission control criteria for the dierent service service class, we describe the notations used in those classes in our model, we rst describe how the worst- conditions: case delays of predictive services are computed. = total link bandwidth, v = utilization target of the link bandwidth 5.1. Worst-case delay computation MIPia (MIPip ) = set of active (passive) ows in MIP class on a link at predictive level i We assume that there are N levels of predictive serp a vices. The delay bound of level i is smaller than the MIPi = MIPi [ MIPi delay bound of level j for i < j . The aggregate token MIGa (MIGp ) = set of active (passive) ows in MIG class on a link bucket lter parameter of predictive class j is (rj ; bj ) P N and is the link capacity such that j=1 rj . MIG = MIGa [ MIGp Parekh, in [11], has shown that in a network with arbi- MDPia (MDPip ) = set of active (passive) ows in trary topology, under WFQ scheduling discipline the deMDP class on a link at predictive level i lay bounds of a guaranteed ow depends only on its re- MDP = MDP a [ MDP p i i i served ow rate and its bucket depth. This implies that, a = MIP a [ MDP a = set of active ows of pre P i i i WFQ isolates the guaranteed ows from each other and dictive level i in addition it isolates all guaranteed ows from all pre C ; ^C ; bC ; ^bC = total aggregate bandwidth, meadictive trac. sured aggregate bandwidth, aggregate bucket depth, When the packet size is small, i.e. the transmission measured aggregate bucket depth of all ows in set time of each packet is suciently small compared to C respectively other delays and hence can be ignored, the worst-case delay, Dj , of priority queue level j is given by the fol- b ; ^b ; r ; r^ = bucket depth, measured bucket depth, lowing theorem in [11]: requested trac rate, measured trac rate of ow Theorem 1 Parekh [11]: The worst-case class j derespectively lay, with FIFO discipline within the class and assuming Jamin et. al.[7] observe that, when a link approaches in nite peak rates for the sources, is full utilization, the variance of the packet delay diverges. In a system using a measurement-based approach the Xj b delay variations may be extremely large when the link i utilization becomes very high. Hence, the admission Dj = i=1j?1 control algorithms are designed to keep the link utiX ? ri lization below a utilization target v (0 v 1). The i=1 appropriate utilization target for a link will depend on for each class j. Further, this delay is achieved for a the characteristics of the trac owing through it (see strict priority service discipline under which class j has Subsection 3.2 in [7]). the least priority. When a mobile host h requests to initiate a ow , 5.2. Admission Control Criteria for MIG service it speci es its trac characteristics with a token bucket The ow , active or passive, is admitted to the MIG lter (r ; b ) and mobility speci cation with a set of service class, if all of the following conditions are satiscells Mh . In addition it speci es the class of service and if it requests MIP or MDP service, it also speci es ed at the switch or base station: the desired delay bound. Depending on the requested 1. After admitting the ow , if all ows of class MIG service class, the switches and the base stations along who have made reservations on this link, both acthe paths from the sender to each of the cell in Mh check tive and passive, becomes active on this link, their appropriate admission control conditions. The new ow worst case bandwidth demands must be satis ed. will be admitted only if the admission control criteria Thus, the total reserved bandwidth of all ows of are satis ed at all of those switches and base stations. MIG class, both active and passive, including r Before describing the admission control criteria for each should not exceed the targeted utilization of the
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
link bandwidth: v > r + MIG
(b)
2. After admitting the ow , if all ows of class MIG and MIP who have made reservations on this link, both active and passive, becomes active on this link, their current measured bandwidth demands must be satis ed. Thus the sum of the requested ow rate r and the current measured rates of the ows in classes MIG and MIP, both active and passive, should not exceed the targeted link utilization:
v > r + ^MIG +
N X ^
MIPi
i=1
3. After admitting the ow , the current measured bandwidth demands of all active ows on the link must be satis ed. Thus the sum of the requested
ow rate r and the current usage of the link should not exceed the targeted link utilization:
v > r + ^MIG
N X a+ ^ a i=1
X
2MIPip
^b is replaced by
X
2MIPip
7
b
All other quantities in the above equation can be estimated by measurement. Therefore, the condition to be satis ed is: DMIP j > DMIPj for 1 j N . 5. The delay bound DP j of the predictive level j; (1 j N ), should not be violated when this new ow is admitted. The new delay of predictive level j , DPja , after admitting is computed by taking into account only the active ows of all classes and is given by the following equation: 0
0
DPja =
Xj ^b a Pi i?1 j ?1
0
The quantity
?^MIGa ?
Xj ^b i?1
X ^ a ? r i=1
Pi
Pi can be computed from the ob-
served delays of the predictive class trac. Therefore the condition to be satis ed is DP j > DPja for 1 j N . Note that, in the above conditions 4 and 5, DP j = DMIP j for 1 j N .
Pi
4. After the ow is admitted, the delay bound of (1 j N ), should not be MIP level j , DMIP j violated when all ows of MIG and MIP class which have made reservations on this link are in active state. The new delay of MIP level j , DMIPj , after 5.3. Admission Control Criteria for MIP service is admitted, is given by: 0
0
Xj bMIP i i DMIPj j? X ?MIG ? MIPi ? r i Xj X b X b ^
0
=1
=
1
^
=
^
=1
(
2MIPia
^ +
i=1 j ?1
?^MIG ?
X X (
i=1 2MIPia
r^
2MIPip
+
^ )
X
2MIPip
r^ ) ? r
Although we can estimate the bucket depth in the equivalent token bucket lter of the aggregate trac of each predictive class from the observed packet delay for that class, it is dicult to estimate the bucket depth in the equivalent token bucket lter of each individual ow in that class. Therefore in the above equation we make the following substitutions: X ^b is replaced by (a) 2MIPia
MIN (
X
2MIPia
b ;
X ^b )
2Pia
The ow , active or passive, will be admitted to the MIP class k at a switch or base station if all of the following conditions are satis ed: 1. This condition is same as the condition 2 in Subsection 5.2. The sum of the requested ow rate r and the current measured rates of the ows in classes MIG and MIP, both active and passive, should not exceed the targeted link utilization:
v > r + ^MIG +
N X ^ i=1
MIPi
2. This condition is same as the condition 3 in Subsection 5.2. The sum of the ow's requested rate r and the current usage of the link should not exceed the targeted link utilization:
v > r + ^MIGa +
N X ^ a i=1
Pi
3. After is admitted, the delay bound DMIP j of the MIP level j , j k should not be violated when all
ows of MIG and MIP class who have made reservation on this link are in active state. The delay
8
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
bound of the MIP class k, DMIPk , is estimated by the following equation after is admitted: 0
Xk ^b
MIPi + b
i=1
DMIPk =
X ^ ?MIG ?
0
k?1
^
Xk ( X
i=1 2MIPia
X
2MIPip
^b ) + b
k? X X ?MIG ? ^
=
DPja = 0
MIPi
i=1
^b +
The new delay of the predictive level j, k < j N , after admitting is estimated by the following equation (considering only active ows):
1
i=1 2MIPi
Xj ^b a + b i=1
?^MIGa ?
Pi
j? X ^ a ? r 1
i=1
Pi
Therefore the condition to be satis ed is: DP j > DPja , for k j N . 0
r^
5.4. Admission Control Criteria for MDP service
For k < j N , the estimated delay, after is The ow will be admitted to the MDP class k on a admitted, is given by : link in an active state, if all of the following conditions Xj ^b + b are satis ed: MIPi 1. This condition is same as the condition 3 in Subi=1 DMIPj = = j ?1 X section 5.2. The sum of the requested ow rate r ^MIPi ? r ?^MIG ? and the current usage of the link bandwidth should i=1 j not exceed the targeted link utilization: X( X ^b + X ^b ) + b N X p a a ^Pia + v > r + ^ i=1 2MIPi MIG 2MIPi 0
j? X X ?MIG ? 1
r^ ? r
i=1
2. The delay bounds of the predictive levels k and above should not be violated after admitting this In the above equations we make the following sub ow, i.e. this condition is same as the condition 4 stitutions: in the previous subsection. X ^b is replaced by The ow will be admitted to the MDP class k on a (a) a 2MIPi X link in a passive state without any admission control; it X ^b ) will only be entered into the set of MDP ows who are MIN ( b ; 2MIPia 2Pia allowed to use the link. X X ^b is replaced by (b) b 5.5. Handling Handos p 2MIP p 2MIP ^
i=1 2MIPi
i
i
Therefore the condition to be satis ed is: DMIP j > DMIPj , for k j N . 0
4. The delay bound DP j of the predictive level j , j k should not be violated when the new ow is admitted. The new delay of the predictive level k after admitting is estimated by the following equation (considering only active ows):
DPka =
Xk ^b a + b i=1
0
?^MIG
DPka +
Pi
= X a? ^ a k?1
i=1 b k?1
?^MIGa ?
Pi
X ^ a i=1
Pi
When a mobile host with an existing ow in a particular service class moves into a new cell within its mobility speci cation, the delay bounds of the already existing ows in the new cell may be violated or the link utilization target may be exceeded. Since the ows of MIG and MIP service classes cannot tolerate the consequences of mobility, we degrade ows of MDP class. There are other alternative ways to deal with hando which are out of scope of this paper. In this paper we consider a simple hando scheme: when a mobile host with a ow of MDP class enters a new cell within its mobility speci cation which cannot provide its existing QoS, the ow is dropped; when a ow of class MIG or MIP moves into a cell, sucient number of MDP ows are dropped to restore QoS bounds.
9
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
6. Reservation Protocol
C9 C8
C10
In ISPN, a reservation protocol, RSV P [18], has been proposed that sets up reservation states at each router along the path from a sender to a receiver. It uses two types of messages to setup the reservation states in the routers, PATH message to setup the data
ow path, and RESV message to make the bandwidth reservation. To accommodate mobile hosts, we propose a reservation protocol, MRSV P , which is an extension of RSV P . It has the following characteristics : 1) It makes an active reservation from the sender to the current base station of the mobile host along which data
ow will start immediately. 2) It also makes passive reservations along the data paths from the sender to all other base stations in the mobility speci cation of the mobile host. We brie y outline the protocol MRSV P . The details of MRSV P can be found in [16]. Let Mh be the mobility speci cation of a mobile host h. A ow may be an unicast ow or a multicast ow. For a multicast
ow, h will join the multicast group. We assume that the underlying routing protocol is the IETF Mobile-IP routing protocol[12], in which data packets are routed to a mobile host via mobility agents, home agent and foreign agent. The protocol works as follows: 1. The sender sends a PATH message to h or the multicast group address. 2. On receiving a PATH message for the rst time, h will send the ow speci cation to all mobility agents in Mh. If the reservation is for a unicast ow, h will send Mh to its home agent. If the reservation is for a multicast ow, h will ask the mobility agents of Mh to join the multicast group. Then it will send a RESV message to the sender to make an active reservation. On receiving subsequent PATH messages, it will only send the RESV message to the sender. 3. On receiving Mh, the home agent will tunnel all PATH messages for h to all mobility agents in Mh . 4. On receiving a PATH message for a non-local mobile host, a mobility agent will send a RESV message to the sender to make a passive reservation. 5. When h detects a signi cant change in the data ow rate, it will send the new ow speci cation to all mobility agents in Mh .
C7 C1
N
C6 C2
C5
C3 C4
Figure 1. Network model for simulation
7. Performance Evaluation We have used simulation experiments to evaluate the performance of the service model, the admission control scheme and the hando policy described above. The main objective of these experiments was to determine the viability of providing mobility independent service guarantees to the ows of mobile users. In our system we can consider three cases: 1. Case I : The network supports only ows of mobility independent service classes. 2. Case II : The network supports ows from both mobility independent and dependent service classes and allows multiplexing of the network resources among them. Let be the fraction of the arriving
ows belonging to the mobility dependent service class and 1 ? of all the arriving ows belong to the mobility independent service classes. 3. Case III : The network supports only ows from mobility dependent service classes. Obviously, in Case I the network utilization will be very low and a very small number of ows can be supported, although no ow will be dropped. On the other hand, in Case III the network utilization and the number of ows supported will be high but a substantial fraction of the ows may suer degradation in QoS or may be dropped when users move from one location to another. By allowing multiplexing of resources among the mobility independent and dependent service classes,
10
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance 1
1.2 MIXED MD MI
MIXED_drop_rate MIXED_mdp_drop MD 1
0.8
0.8 0.6 0.6 0.4 0.4
0.2 0.2
0
0 0
0.5
1
1.5
2 2.5 3 flow arrival rate (flows/sec)
3.5
4
4.5
5
Figure 2. Link utilization Vs. ow arrival rate (homogeneous)
0
0.5
1
1.5
2 2.5 3 flow arrival rate (flows/sec)
3.5
4
4.5
5
Figure 3. Drop rates Vs. ow arrival rate (homogeneous)
#flows
200 the network utilization and the number of ows supMIXED MD ported can be increased from those values in Case I. MI From the simulation experiments, we would like to see the eects of multiplexing ows from the mobility inde150 pendent and dependent service classes on the network resource utilization, number of ows supported and the fraction of ows dropped. If we can nd that in Case 100 II, by multiplexing ows of mobility independent and dependent service classes the network utilization and the number of ows supported can be kept high (ide50 ally close to those values in Case III) and the fraction of ows dropped is not substantially greater than those in Case III, we can claim that it is viable to support mobility independent services provided sucient mul0 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 tiplexing of resources among mobility independent and flow arrival rate (flows/sec) dependent service classes are allowed. As a by product, we can also determine the values of , the degree of Figure 4. Active ow counts Vs. ow arrival rate (homogeneous) multiplexing, for which we can obtain this performance. ure 1). In this con guration, the network has a star topology. There are ten wireless cells, C1-C10, each cov7.1. Simulation Model ered by a base station. The cells are non-overlapping There are many experiments in the literature that and are arranged in a ring con guration. There is a have investigated the impact of various admission con- central node N to which each of these base stations is trol policies on the network performance and QoS pa- connected by a link of in nite bandwidth. The wireless rameters of ows. These results have considered the ef- link at each cell has a bandwidth of 2Mbps. A mobile fects of ow characteristics and network con gurations node can move to any of its two neighboring cells on the on those performance parameters. In our simulation ring. Each ow originates at the central node and the experiments we wanted to determine the impact of mo- data ow downstream to a mobile host in a cell via the bility of hosts on the dierent performance parameters. base station of that cell. We considered only unicast Therefore we used a simple network con guration (Fig- ows.
11
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance 200
1 total MIG MIP MDP
utilization mdp_drop drop 0.8
150
#flows
0.6 100 0.4
50 0.2
0
0 0
0.5
1
1.5
2 2.5 3 flow arrival rate (flows/sec)
3.5
4
4.5
5
Figure 5. Active ow counts in MIXED Vs. ow arrival rate (homogeneous)
0
0.6
0.8
1
Figure 6. Link utilization, drop rate and mdp drop rate Vs. mdp frac (homogeneous) 200 total MIG MIP MDP
The main performance parameters we measured are link utilization, ow drop rate and active ow count. These parameters are de ned as follows:
150
#flows
link bandwidth used. The average link utilization is the average of the utilization of the link over a period of time. ow drop rate : It is de ned as the fraction of the number of ows which were admitted but later dropped over all links in the network due to violation of bandwidth or delay constraints when users move from one location to another. active ow count : It is the count of the number of active ows on all links of the network. The average active ow count is the average of the number of active ows on all links over a period of time.
0.4 mdp_frac
7.2. Parameters
link utilization : It is de ned as the fraction of the
0.2
100
50
0 0
0.2
0.4
0.6
0.8
1
mdp_frac
Figure 7. Active ow counts Vs. mdp frac (homogeneous)
We also measured the average active ow count for each service classes. These are de ned as follows: av active count = average number of active ows We measured two types of ow drop rate: drop rate and mdp drop rate. These are de ned as: of all classes. drop rate = d =a1 mig av active count = average number of active mdp drop rate = d =a2
ows in the class MIG. where mip av active count = average number of active d = number of ows dropped.
ows in the class MIP. a1 = number of ows admitted to all classes. mdp av active count = average number of active a2 = number of ows admitted to the class MDP. ows in the class MDP.
12
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
There are many parameters associated with the ow and the host mobility. We took measurement data for dierent values of these parameters. These parameters are: ow arrival rate : The rate at which new ows arrive. ow duration : Duration of a ow. ow spec : Data trac generation characteristics of a ow. This is characterized by the token bucket lter parameter of the ow. cell stay time : The period of time after which a mobile host moves into a neighboring cell. mobility spec : The length of the ring over which a host can move. This is expressed in number of adjacent cells. The system parameters which we considered are: mdp frac : Fraction of the arriving ows belonging to the class MDP. mig frac : Fraction of the arriving ows belonging to the class MIG. mip frac : Fraction of the arriving ows belonging to the class MIP. It should be noted that: mdp frac + mig frac + mip frac = 1 . 7.3. Simulation Experiments
1 MIXED_mdp_drop MIXED_drop MD 0.8
0.6
0.4
0.2
0 0
0.5
1
1.5
2 2.5 3 flow_arrival_rate (flows/sec)
3.5
4
4.5
5
Figure 9. Drop rates Vs. ow arrival rate (non-homogeneous)
We performed several simulation experiments to measure the performance parameters by varying dierent sets of ow, mobility and system parameters. Our goal was to investigate the viability of multiplexing network resources among the ows of mobility independent and dependent service classes. The degree of multiplexing could be controlled by varying the values of the system parameters mdp frac, mig frac and mip frac. The two extreme cases are when mdp frac = 1 (i.e. all ows belong to mobility dependent class) and mdp frac = 0 (i.e. when there are no mobility dependent ow).
1 MIXED MI MD
200 MIXED MI MD
0.8
150
#flows
0.6
0.4
0.2
100
50
0 0
0.5
1
1.5
2 2.5 3 flow_arrival_rate (flows/sec)
3.5
4
4.5
5 0 0
Figure 8.
Link utilization Vs. ow arrival rate (nonhomogeneous)
0.5
1
1.5
2 2.5 3 flow_arrival_rate (flows/sec)
3.5
4
4.5
5
Figure 10. Active ow counts Vs. ow arrival rate (nonhomogeneous)
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
individual ows. Therefore, in this paper, we did not consider any speci c trac model for the data sources in the simulation experiments. As a consequence, the measured trac rate of a ow was always equal to its requested trac rate. We performed two types of experiments: homogeneous, i.e. when the ows of mobility independent and dependent classes have the same ow and mobility characteristics, and non-homogeneous, i.e. when the ows of mobility independent and dependent service classes have dierent ow and mobility characteristics.
200 total MIG MIP MDP
150
#flows
13
100
50
7.4. Simulation Results 0 0
0.5
1
1.5
2 2.5 3 flow_arrival_rate (flows/sec)
3.5
4
4.5
5
Figure 11. Active ow counts in MIXED Vs. ow arrival rate (non-homogeneous)
We have chosen the values of dierent ow and mobility parameters as follows. We assumed poisson arrival of ows i.e. interarrival times between successive
ows is exponentially distributed. Flow duration and cell stay time are also exponentially distributed. We categorized the incoming ows into three classes, F1, F2 and F3, each class having a xed token bucket lter parameter and delay bound requirements. Incoming
ows are almost equally distributed among these three classes. Incoming ows were assigned to the three service classes MIG, MIP and MDP according to the values of mdp frac, mig frac and mip frac. In the following we describe the notations used for ow and mobility parameters:
a : Mean ow arrival rate in number of ows per
We performed a large number of simulations with dierent sets of values for dierent ow, mobility and system parameters. In both homogeneous and nonhomogeneous experiments, we performed two sets of simulations: in the rst set, called Set-I, we measured the performance parameters by varying the mean ow arrival rate and keeping all other ow, mobility and system parameters xed; in the second set, called Set-II, we measured the performance parameters by varying the class fractions mdp frac, mig frac and mip frac and keeping all other ow and mobility parameters xed. The Set-I experiments consists of three dierent subsets of experiments corresponding to three dierent sets of system parameter values. Two of them are for the two extreme cases of multiplexing: in the rst one we set mdp frac = 1 i.e. all ows belong to the mobility dependent service class, we call this MD; in the second one, we set mdp frac = 0, i.e. all ows belong to the mobility independent service class (either MIG or MIP), we call this MI. In the third subset, we used a non-zero value of mdp frac, i.e. there are ows from both mobility dependent and independent service classes, we call this MIXED. In the following we describe the chosen values of the parameters and the simulation results for the homogeneous and non-homogeneous experiments.
second dl : Mean duration of ows for longer duration ows ds : Mean duration of ows for shorter duration ows hm : Mean mobility rate in number of moves per second for highly mobile hosts. lm : Mean mobility rate in number of moves per 7.4.1. Homogeneous In the Set-I experiments we used the following values second for mobile hosts with low mobility. Then, mean cell stay time = 1/mean mobility rate. for the ow, mobility and system parameters which were s : Mean mobility spec in number of adjacent cells. xed: The variation in the bandwidth demands on a link due dl = ds = d = 300 seconds to mobility of hosts is signi cantly more pronounced hm = lm = m = 0.2 than the variation of the bandwidth requirements of s = 5
14
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance 1 utilization mdp_drop drop_rate 0.8
0.6
0.4
mip frac = 0:25
flow spec:
F1 : token bucket lter = (64 kb/s, 4 kb), delay = 50 ms F2 : token bucket lter = (128 kb/s, 10 kb), delay = 100 ms F3 : token bucket lter = (200 kb/s, 20 kb), delay = 500 ms
The simulation results are shown in the following Figures 2-5. In Figure 2 we compared the network link utilization among the MD, MI and MIXED experiments. 0 Utilization level is much lower in MI, because all ows 0 0.2 0.4 0.6 0.8 1 mdp_frac make passive reservations at all locations where they Figure 12. Link utilization, mdp drop rate and drop rate Vs. may move. As expected, utilization is highest in MD because all ows make reservations at its current locamdp frac (non-homogeneous) tion only, and hence a large number of ows can be 200 admitted. For the MIXED experiments utilization level total MIG is very close to that of the MD experiments. This imMIP MDP plies that, network utilization does not suer signi cantly when we allow multiplexing of mobility indepen150 dent and dependent ows in the same network. In Figure 3, we compare the drop rates of the MD and MIXED experiments. We observe that the drop rates of 100 the MDP ows (mdp drop rate) in the two experiments are almost same. Whereas, if we consider overall drop rates (drop rate), it is lower in MIXED than the drop 50 rate in MD. In Figure 4, we compare the average number of active ow counts in MD, MI and MIXED experiments. Obviously, in the case of MI it is much lower than that 0 0 0.2 0.4 0.6 0.8 1 in MD. However, there is no signi cant dierence bemdp_frac tween the average number of active ows in MD and Figure 13. Active ow counts Vs. mdp frac (non-homogeneous) MIXED experiments. In Figure 5, we observe that, a signi cant fraction of the ows in MIXED experiments For the MD experiments: belong to the classes MIG and MIP. mdp frac = 1:0 In the experiments in Set-II, we kept the ow armig frac = 0:0 rival rate xed at a = 2:0 and varied the value mip frac = 0:0 of mdp frac from 0 to 1 (mig frac = mip frac = (1For the MI experiments: mdp frac)/2). In Figure 6, we observe that, the netmdp frac = 0:0 work utilization increases as mdp frac increases. When mig frac = 0:5 mdp frac > 0:3, utilization reaches within 20% of the mip frac = 0:5 utilization achieved when mdp frac = 1. From Figure For the MIXED experiments: 7, we see that when mdp frac is in the range (0.3, 1.0), mdp frac = 0:5 av act count and mdp drop rate remains almost same. mig frac = 0:25 Also, when mdp frac is in the range (0.3, 0.8), a signif#flows
0.2
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
icant fraction of the active ows belong to the classes MIG and MIP. From these results we conclude that, the performance of the system does not degrade much when we allow multiplexing of mobility dependent and independent
ows in the same network. Thus it is feasible to provide mobility independent service in a network when multiplexing of mobility dependent and independent ows are allowed suciently.
15
The simulation results for the experiments of Set-I are shown in Figures 8-11. In Figure 8-9, we see that when the network is saturated with demands, the utilization in MIXED experiments is within about 10% of the utilization in MD experiments, but the ow drop rate in MIXED is less than the drop rate in MD experiments. In Figures 10-11, we observe that, when the network is saturated, a slightly larger number of ows are supported in MIXED experiments than in the MD experiments. 7.4.2. Non-homogeneous In the Set-II experiments, the ow arrival rate was In these experiments we used the following set of xed at a = 2:0 and the value of mdp frac was varied values for the dierent parameters. from 0 to 1 (mig frac = mip frac = (1-mdp frac)/2). The simulation results are shown in Figures 12 and 13. dl = 1000 seconds, ds = 200 seconds The performance characteristics of the system is similar hm = lm = m = 0.01 to those obtained in homogeneous experiments. s=5 From these results we observe that, if sucient multiplexing of the ows of mobility dependent and inde For the MD experiments: pendent ows are allowed and the ows with longer dumdp frac = 1:0 ration are provided with mobility independent services, mig frac = 0:0 the drop rates of mobility dependent ows decreases mip frac = 0:0 and more ows can be supported without signi cant For the MI experiments: degradation of utilization. mdp frac = 0:0 mig frac = 0:5 mip frac = 0:5 8. Conclusion and Future Work For the MIXED experiments: mdp frac = 0:5 In this paper we have investigated the problems of mig frac = 0:25 providing real-time service guarantees to the mobile mip frac = 0:25 hosts in an Integrated Services Packet Network. We presented a service model for real-time services to mo flow spec: bile hosts that can accommodate highly adaptive appliF1 : token bucket lter = (64 kb/s, 4 kb), cations as well as applications which require mobility delay = 50 ms independent QoS guarantees. We proposed an admisF2 : token bucket lter = (128 kb/s, 10 kb), sion control scheme for this model that can achieve very delay = 100 ms high utilization of network bandwidth. F3 : token bucket lter = (256 kb/s, 20 kb), We evaluated the performance of our system by simdelay = 500 ms ulation. Simulation results show that, when sucient degree of multiplexing of mobility independent and deIn both MI and MD experiments, 50% of the incom- pendent classes are allowed, network performance does ing ows had a mean duration dl and the other 50% not degrade signi cantly and for certain ow and mobilof the ows had mean duration ds . In MIXED exper- ity characteristics, ow drop rate and number of ows iments, ows of the classes MIG and MIP had mean supported can be improved over the model in which duration dl and the ows of the class MDP had mean only mobility dependent services are provided. duration ds . Thus we could compare the results of the We expect that, the network performance will dethree experiments when the ows had similar charac- pend on the user mobility patterns. Several mobility teristics but they diered in the service guarantees they models have been described in a previous work[13]. As received. part of future work, we plan to investigate the impact of
16
Talukdar et. al. / Integrated Services Packet Networks with Mobile Hosts: Architecture and Performance
user mobility on the network performance by considerNetwork, In Proceedings of The INFOCOM 1997, Kobe, Japan, April 1997. ing various user mobility models. Also, we are currently [16] Talukdar, A. K., Badrinath, B. R. and Acharya, A., looking at various issues of hando management techMRSVP: A Reservation Protocol for an Integrated Services niques. Packet Network with Mobile Hosts, Dept. of Computer Sci-
References [1] Acampora, A. S. and Naghshineh, M., An Architecture and methodology for Mobile-Executed Hando in Cellular ATM Networks., IEEE JSAC, Vol 12, No. 8, October 1994. [2] Badrinath, B. R. and Talukdar, A. K., IPv6 + Mobile-IP + MRSVP = Internet Cellular Phone ?, In Proceedings of the International Workshop On Quality of Service, May 1997. [3] Bakre, A. and Badrinath, B. R., Hando and system support for indirect TCP/IP, In Proc. of the 2nd Usenix symposium on mobile and location independent computing, pp. 11-24, April 1995. [4] Balakrishnan, Hari, et.al., A comparison of mechanisms for improving TCP performance over wireless links, In Proc. SIGCOMM'96, August 1996. [5] Clark, D.D., Shenker, S. and Zhang, L., Supporting RealTime Applications In An Integrated Services Packet Network: Architecture and Mechanism., Proc. SIGCOMM '92, 1992. [6] Demers, A., Keshav, S. and Shenker, S., Analysis and simulation of a Fair Queuing Algorithm, Proc. SIGCOMM '89, Sept. 1989 [7] Jamin, S., Danzig, P.B., Shenker, S and Zhang, L., A Measurement-based Admission Control Algorithm for Integrated Services Packet Networks., Proc. SIGCOMM '95, 1995. [8] Lee, K., Adaptive Network Support for Mobile Multimedia, In Proc. of the 1st Annual International Conference on Mobile Computing and Networking, pp. 62-74, November 1995. [9] Levine, D. A., Akyldiz, I. F. and Naghshineh, M., The Shadow Cluster Concept for Resource Allocation and Call Admission in ATM-Based Wireless Networks, In Proc. of the 1st Annual International Conference on Mobile Computing and Networking, pp. 62-74, November 1995. [10] Lu, S. and Bharghavan, V., Adaptive Resource Management Algorithms for Indoor Mobile Computing Environments, Proc. SIGCOMM'96, August 1996. [11] Parekh, A. K., A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks., PhD Theses, MIT, LIDS, Tech. Report LIDS-TR-2089 1992. [12] Perkins, Charlie, IP Mobility Support, IETF Draft, May 1996. [13] Rajagopalan, S. and Badrinath, B. R., An Adaptive Location Management Strategy for Mobile IP, In Proc. of the rst Annual International Conference on Mobile Computing and Networking, November 1995. [14] Singh, S., Quality of Service guarantees in Mobile Computing, Journal of Computer Communications, vol. 19, 1996. [15] Talukdar, A. K., Badrinath, B. R. and Acharya, A., On Accommodating Mobile Hosts in an Integrated Services Packet
ence Technical Report, DCS-TR-337, Rutgers University, USA. [17] Turner, J., New Directions in Communications, or Which way to the Information Age?, IEEE Communications Mag., vol 24, pp 8-15, 1986. [18] Zhang, L., Deering. S., Estrin. D., Shenker, S. and Zappala, D., RSVP: A New Resource ReSerVation Protocol. IEEE Network Sept. 1993.