Fashion & Beauty

A two-level resource management scheme in wireless networks based on user-satisfaction

A two-level resource management scheme in wireless networks based on user-satisfaction
of 11
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
   A Two-level Resource Management Scheme inWireless Networks Based on User-Satisfaction  ∗ Sourav Pal a Mainak Chatterjee b Sajal K. Das a a Department of Computer Science and Engineering, University of Texas at Arlington, TX, USA b Department of Electrical and Computer Engineering, University of Central Florida, FL, USA The success of future generation wireless data services will depend on the parameterized  provisioning of quality of service (QoS) for applications whose demands and nature arehighly heterogeneous. Also, user satisfaction will play a key role in the economic viabilityof wireless service deployments. In this paper, we present a QoS framework based on the paradigm of   traffic class  and   user satisfaction . We address the problem of dealing with sub- jectiveness of user satisfaction or expectation from service providers by defining what wecall  user irritation factors  , using Sigmoid functions. These factors reflect users’ sensitivityand tolerance to delay. The proposed class-based QoS framework comprises a radio re-source management scheme which considers user satisfaction based on the perceived QoS,and caters to heterogeneous applications that have diverse QoS requirements. Our resourcemanagement scheme has two components: the admission control algorithm caters to the long term  user satisfaction while the session-based rate and bandwidth allocation schememanipulates the  short term  user satisfaction. Soft-reservation schemes are also proposed to cater to the higher paying users. Performance metrics have been specifically defined  for each traffic class. Extensive simulations using four types of traffic and three classes of users reveal that the proposed framework offers improved QoS without compromising theutilization of the system. I. Introduction The demand for wireless data services has led to theevolution of third generation ( 3 G ) wireless serviceswhich deliver a broad range of multimedia applica-tions to mobile users. The transition from traditionalvoice services to data services with heterogeneousrequirements necessitates a revisit of the radio re-source management schemes. Resource managementmust consider the impact of error prone transmissionmedium, heterogeneity of application requirements,and issues related to fairness among users. Also, thereis a need to differentiate users based on the amount of revenue they are willing to pay and their expectationsfrom the services.We envision that the success of wireless data ser-vices in conjunction with traditional voice serviceswould ultimately depend on  user satisfaction . Thusa QoS framework needs to be developed that identi-fies user satisfaction and also facilitates negotiationbetween the users and the service providers. Iden-tifying the relevant QoS for each of the diverse ser-vices and distinguishing the variation of user satisfac- ∗ This work is partially supported by NSF ITR grant IIS-0326505. tion with the perceived QoS is an important researchchallenge. User satisfaction depends on the subjectiveexpectation of the service and hence varies with thetype of services.Several resource management schemes have beenproposed which address the various requirementslike heterogeneous service demands, fairness, star-vation, channel transmission error and delay bounds(for example see [16] and references therein). Theyinclude power control algorithms at the physicallayer, scheduling or rate-adaptation algorithms at themedium access control (MAC) layer and service ad-mission algorithms at the network layer. However,many such schemes prioritize the voice services andallocate the residual bandwidth to non-real time appli-cations which are deprived of any assurance on the de-lay bound. A user may care less about the per-packetdelay and might put more emphasis on the downloadtime of the  entire  data. This motivates us develop re-source management schemes which strives topreservethe QoS requirements of different heterogeneous ser-vices, while maintaining fairness amongst differentclasses of users.In this paper, we propose a radio resource manage-ment framework which tries to adhere to the QoS re-4  Mobile Computing and Communications Review, Volume 9, Number 4  quirements of the applications by exploiting the  sub- jectiveness  associated with user satisfaction. We cate-gorize users intodifferent classes based on therevenuepaid, and consider that all the classes are endowedwith heterogeneous wireless services. We propose thenotions of   short term irritation  and  long term irrita-tion , extend them to multiple traffic classes and pro-pose service (call) admission algorithms and schedul-ing policies. These policies are tailored for UniversalMobile Telecommunications System (UMTS) definedtraffic. Priorities are also considered among theseclasses. The proposed two-level resource manage-ment scheme try to improve the delay in deliveringthe entire non-real-time content and real-time traffic(conversational/streaming) by taking into considera-tion the short- and long-term effects of user irritation.More specifically, the proposed call admission con-trol algorithm regulates the long-term irritation of theusers, whereas the short-term satisfaction of the usersis guaranteed by the scheduling policy. It not onlyprovides a bound on the delay but also manipulatesthe resources so as to maintain the irritation of eachuser below a certain threshold.The remainder of the paper is organized as follows.The network architecture along with how the user ser-vice level agreement and the policy management is-sues are stored in the databases are discussed in sec-tion II. The various traffic classes and the right metricfor QoS characterization are presented in section III.The subjective user satisfaction model and its depen-dence on different service types and perceived QoS isstudied in section IV. The two-level radio resourcemanagement scheme is proposed in section V whilesection VI presents the simulation model and the ex-perimental results. Conclusions are drawn in the lastsection. II. Architecture and Policies In this section, we discuss the network architecturealong with the databases which contain the user pro-files. These databases also hold the policies pertainingto the user classes and their respective quality of ser-vice expectations from the system. II.A. Network Architecture The resource management framework on which ourproposed scheduling algorithm operates is based onthe architecture components proposed in the IETFPolicy Information Base (PIB) [6] for differentiatedservices. Figure 1 shows the architectural overview of a 3G wireless cellular network that consists of three PDSNP D P AirLink  Internet TerminalBTS LDAPSLA DB Figure 1: Network Architectureimportant network elements : (1) the airlink and ter-minal, (2) the base station transceiver system (BTS)and (3) packet data service node (PDSN). Each cellexecutes a unique copy of the proposed scheduling al-gorithm for handling requests generated in that cell.In the 3G system, the PDSN includes the gatewayfunctions that interconnect the Internet domain. ThePIB framework includes components like (1) ServiceLevel Agreement (SLA) database, (2) policy decisionpoint (PDP), and (3) policy execution point (PEP).The air link supports  uplink   and  downlink   channels.The uplink channel transmits the request of the clientsto the server, where the scheduler schedules the datato the clients through the downlink channels. Thoughthe uplink bandwidth is smaller, we assume that thereexist no uplink channel contention between differentclients sending requests to the server. Location depen-dent channel transmission errors which are bursty innature needs to be considered for the scheduler designto achieve accuracy and efficiency of the system. II.B. PolicyManagementandSLAIssues It is to be noted that many wireless carriers are alreadyadopting whatcanbe called  differentiated service con-tracts  for voice services to mobile users. However, nosimilar effort has been made for data services. The ob- jective here is to create different classes of customersbased on the selected service packages for data ser-vices. The distinction in QoS levels lies in the band-width provided and hence the delay and throughputoffered to the different classes of users. The policymanagement and the customer service agreements arestored in the SLA database as a set of rules. The pol-icy rules are created using the IntServ/DiffServ QPIM(quality policy information modeling) technique as re-ported in most IETF drafts, for example, in [13]. ThePDP contains all the PIBs [6] in addition to the MIBs(management information bases) required for policymanagement.The PDP function for the differentiated  Mobile Computing and Communications Review, Volume 9, Number 4  5  services can be located in the PDSN or mobile switch-ing center(MSC). The policy execution function inthis case remains within the radio network controller(RNC). We consider three service classes -  Gold, Sil-ver  , and  Bronze , where each user class supports allthe services. We propose that the PIB support allthree user classes but with different commitment lev-els. The classification being dependent on the QoSthey expect from the network and the revenue they arewilling to pay. The Gold class pays the highest rev-enue and the Bronze class pays the least. The delaysuffered for the same service is thus highest for anBronze class client and least for a Gold Class client.The fairness of service in this context is relative tothe price paid by the clients. The scheduler inter-nally classifies the users on the basis of the  user ir-ritation factor (UIF) , assigning higher UIF to a userwith higher priority. This is modeled in the next sec-tion.The PIB also specifies that service to client requestswould be processed in the following manner.  Guaran-teed QoS   mode of service implies that the system willbe able to honor the bounded delay.  Negotiable QoS   iswhen the system possesses insufficient or no resourceat all for a request made. However, if the system isoptimistic about being able to serve the request withinthe bounded delay with the anticipation that on-goingtransmissions might release resources in near future,then the request is admitted. Nonetheless, the admit-tance is not strict in nature and the delay restrictionsmight be violated. The third scenario is when the sys-tem is well aware that under no circumstances it canhonor the client request within the specific deadline.It simply rejects the request. III. Traffic Classes and QoS Metrics The proposed QoS framework caters to the diversemultimedia applications as well as the traditionalvoice calls in wireless networks. The framework sup-ports multiple classes of users with different prioritiesand enables fair sharing of the radio resource basedon the user class and subjective satisfaction. More-over, service negotiation between users and serviceproviders as in [12] is flexible in the sense that serviceclasses can be pre-configured with the user’s applica-tions, or explicitly selected at the time of initiation of the applications.For the purpose of illustration and simplicity, weconsider the following four QoS classes as proposedfor UMTS networks [1, 2]:  conversational, stream-ing, interactive,  and  background  . However, the pro-posed framework is generic enough and can be ex-tended to any number of traffic classes and services asdesired by the network operator. These heterogeneoustraffic/service classes have specific QoS requirements.The main distinguishing factor is the delay sensitivityof each class. We proceed to outline the different at-tributes of each traffic class and use them to model theuser satisfaction. III.A. Conversation Class This class is mainly intended to be used to carry real--time traffic flows. Time relation (variation) betweeninformation entities of the flow is preserved in conver-sation class. The other fundamental characteristic isthat it has extremely stringent and low delay require-ment. Voice and video telephony are the target appli-cations that falls within the domain of this class.We use  time-hysteresis outage probability  [17] asthe QoS metric for the conversational class. Outageprobability is a classical metric in cellular systemswhich is defined as the probability that the receivedsignal to noise ratio (SINR) will drop below a speci-fied  E  b /N  o , where  E  b  is the energy per bit and  N  o  isthe noise power. The assumption is that the bit rate re-quirement and the bit error rates can be mapped ontoan equivalent  E  b /N  o . III.B. Streaming Class The streaming class is very similar to the conversa-tion class but is less delay sensitive. The other distin-guishing factor is that applications in this class, suchas streaming video, are uni-directional (i.e. one-waytransport). However, the fundamental criterion oftimerelation being preserved between information entitiesremains the same.Though delay and bit error rate affect streamingsessions,  rate jitter   is the most important parameterinfluencing the quality of such traffic. Hence we em-ploy rate jitter as the QoS metric for modeling usersatisfaction, or in other words, for quantifying user ir-ritation for streaming class applications. III.C. Interactive& BackgroundClasses Interactive and background classes are mainly meantto be used by traditional Internet applications likeWWW, Email, Telnet, FTP and News. Since the de-lay requirements of these classes are more slack com-pared to conversational and streaming classes, theyprovide better error rate by means of channel codingand retransmission. The main difference between In-teractive and background classes is that the interac-6  Mobile Computing and Communications Review, Volume 9, Number 4  tive class is mainly used by interactive applicationslike network games and chats, while the backgroundclass is meant for background traffic such as emailsor web downloading. Responsiveness of the inter-active applications is ensured by separating interac-tive and background applications. Traffic in the in-teractive class has higher priority in scheduling thanbackground class traffic, so background applicationsuse transmission resources only when interactive ap-plications do not need them. This is very importantin wireless environment where the bandwidth is lowcompared to wire-line networks.We observe that the delay a user is willing to toler-ate before canceling a session can be considered as asuitable metric for these classes. Though the amountof acceptable delay depends on the particular user, itcan be treated as a tunable parameter which varieswith the user class itself. An additional constraint isthat there should be no data loss. IV. Modeling User Irritation Factor The success of our scheduling algorithm lies in mod-eling the irritation/satisfaction of the user. A basic un-derstanding of the client irritation, i.e., what amountof performance degradation the customer is ready tosuffer without complaining, will enable the sched-uler to estimate the resources (i.e., number of chan-nels) that has to be allocated to a particular request.This will directly help in maintaining the delay boundand indirectly help the service provider to control thechurn factor [8].In the following, we propose a method to model theuser irritation and present two new metrics:  short termuser irritation factor (SUIF)  and  long term user ir-ritation factor (LUIF) . Each factor signifies differentlevels of user satisfaction as described below. Qualita-tively,  SUIF   measures the delay that the user is readyto suffer prior to which the user decides to changeor cancel the particular request.  LUIF   determines thetolerance or irritation of the user resulting from con-tinued degradation of service after which the clientdecides to cancel service completely. For differentclasses as specified in the SLA, the  User IrritationFactor (UIF)  will vary. A high priority user payinghigher revenue expecting lesser delay will be assigneda higher UIF. The goal of the scheduler will be toschedule requests for each client such that the UIF isnot violated.A Sigmoid function has been used in the literatureto approximate the user’s satisfaction with respect toservice qualities or resource allocations [14, 15]. For −50−40−30−20−100102030405000. Utility Parameter    V   a   l   u   e   o   f   U   t   i   l   i   t   y   F   u   n   c   t   i   o   n   α  = 0.1 α  = 0.3 α  = 0.6 α  = 0.9 Figure 2: Examples of Utility Functionsmodeling the satisfaction/dissatisfaction of users, wealso use the Sigmoid function and correlate it with theproposed metrics, SUIF and LUIF. For a random vari-able  x  representing a service parameter like coverageor reliability, the corresponding satisfaction,  U  ( x ) , isgiven by U  ( x ) = 11 + e − α ( x − β  ) .  (1)Here  α  and  β  , determining the steepness and the cen-ter ofthe curve respectively, can betuned tocustomizethe function for different users. The plots for Equation(1), for different values of   α  are shown in Figure 2.From the figure we observe that the satisfaction (util-ity function) increases with increasing  x . But for pa-rameters like price or delay, the satisfaction decreaseswith increasing  x . In such cases we can model satis-faction as U  ( x ) = 1 −  11 + e − α ( x − β  ) .  (2)The value of   α  indicates user’s sensitivity to the QoSdegradation while  β   indicates the “acceptable” regionof operation. We use Equation (2) to model the UIF.In the following section we determine SUIF and LUIFanalytically for each class of traffic. IV.A. Short Term User Irritation Factor The SUIF is measured on a per-session per-user ba-sis. It is also responsible for distinguishing between acall type - new or handoff call and accordingly modelthe user irritation. An in-session user if deprived of service due to handoff, would suffer from greater irri-tation than auser whose request isblocked. Hence, wepropose a simple mechanism to assign τ  1  and τ  2  signi-fying the quantitative factors associated with irritationsuffered due to a new and handoff call, respectively,where  τ  1  < τ  2  <  1 . In the rest of the derivations, weshall use  τ   =  τ  1  or τ  2  depending on the request being  Mobile Computing and Communications Review, Volume 9, Number 4  7  a handoff or a new call. Also, all the random variables x i,j  are normalized with the best possible value being 0  (representing zero delay and jitter) and the worst be-ing  1 . Equation (2) utilizes the  x i,j s defined later tomeasure the SUIFs. SUIF for Class 1:  A classical performance metricin cellular systems is the  outage probability  whichis usually defined as the probability that the QoSprovided to an existing connection will drop belowa certain threshold, or it is the probability that thereceived SINR will drop below a specified  E  b /N  o ,where  E  b  is the energy per bit and  N  o  is the noisepower. The assumption is that the bit rate require-ment and the bit error rates can be mapped onto anequivalent  E  b /N  o . We argue that the SINR translatesto the QoS estimation at the physical layer [9]. Clas-sical definitions of outage probability ( P  op ) based onmarginal statistics fail to capture the true characteri-zation of the dynamism at the physical layer. It hasbeen shown that higher-order statistics of the wire-less channel errors affects the performance of the up-per layers of the protocol stack. Thus a more generaldefinition of outage probability which considers the time  dependencies and durations of the unpredictableevents is needed. To this end, the concept of   time-hysteresis outage probability  [17] is a more relevantperformance metric for voice and data communica-tions. We also feel that time-hysteresis outage proba-bility is a good representation and captures the systemperformance with respect to QoS as well. Thus, wemodel the SUIF for class 1 on time-hysteresis outageprobability as the metric for the system performancein our context. If   x 1 ,  j  denotes the random variablerepresenting the SUIF for class 1 for the  j th user, then x 1 ,  j  =  τ   × P  out,j  (3)where  P  out,j  is the time-hysteresis outage probabil-ity for the  j th user. Also  SUIF  max,V    is defined asthe threshold SUIF crossing which the voice call isdropped. SUIF for Class 2:  We utilize  rate jitter   as the QoS pa-rameter to model the SUIF for streaming traffic. Jitter,quantified in two ways –  delay jitter   and  rate jitter   isintroduced due to variable queuing and propagationdelays. Rate jitter [10] which measures the differencebetween the minimal and maximal inter-arrival packettimes is more appropriate for streaming services thandelay jitter. It actually bounds the difference in packetdelivery rates during the entire period of service forthat particular session and thus is an ideal metric forquantifying user irritation. The higher the rate jitter,the higher the user irritation and vice versa. Thus if  η max,i,j , η min,i,j , ψ i,j  and x 2 ,  j  respectively denote themaximum inter-arrival time, minimum inter-arrival,rate jitter and the random variable representing theSUIF for the streaming class respectively for the  i th session of the  j th user, then x 2 ,  j  =  τ   ×  ψ i,j η max,i,j (4)where  ψ i,j  =  η max,i  − η min,i  is defined as the rate jitter. Again,  SUIF  max,S   is defined as the thresholdbeyond which the call is dropped. SUIF for Classes 3 and 4:  Wedefine the SUIFfor theinteractive and background classes as the delay that anuser is ready to endure before he decides to cancelshis request is a measure of his irritation. Additionalconstraint specific to this class is that there should beabsolutely  no  data loss. The ideal transfer time ( ∆ )for a file of size  S   Kbytes is  ∆ =  S BW   , where  BW  Kbps is the ideal bandwidth supported by the systemfor that user. However, due to congestion, an admit-ted new request might suffer a delay even before it isscheduled for service. The system can afford to as-sign a greater delay ( δ  ext ) to class 4 traffic than class3 since the delay requirements for background trafficis much less strict than interactive. Let the availablebandwidth to the new request be denoted by  BW  r .Hence the actual delay  δ  , suffered by the user is givenby  δ   =  δ  ext  +  S  eff  BW  r where  BW  r  is the bandwidth as-signed in reality to the user and  S  eff   is the effectivedata size [11] that needs to be transmitted due to re-transmission on account of frame error rate (FER).The scheduler is designed to exploit the sensitiv-ity of human nature to delay by transmitting the mainpage (for Web traffic) and some initial data for class4 at the earliest possible time. Scheduling the inter-mediate packets on a regular basis will keep the usersatisfied and also provide the scheduler more time totransmit the entire data. We assume that the maxi-mum delay that any user would be ready to toleratewill be  n  (some multiple) times the ideal time neededto deliver the data, i.e.,  n × ∆ . Thus, if   δ > n ∆ , therequest is serviced in the negotiated mode. We definethe random variable  x ( 3 , 4 ) ,  j  denoting the SUIF of the  j th user for class 3 and 4 traffic as x ( 3 , 4 ) ,  j  =  τ   ×  δ  i,j  − ( n − 1)∆ i,j ∆ i,j (5)where  δ  i,j  and  ∆ i,j  are the ideal and actual delay forthe i th session of the  j th user. The worst case boundeddelay is  δ   =  n ∆ . The corresponding SUIF is termedas SUIF max,BI  .8  Mobile Computing and Communications Review, Volume 9, Number 4
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks