AAA (Allocation, Admission and Assignment Control) for Network management

Noname manuscript No.(will be inserted by the editor)

A Joint Allocation, Assignment and AdmissionControl (AAA) Framework for Next GenerationNetworks

M. V. Ramkumar · Rasmus Hjorth

Nielsen · Andrei Lucian Stefan · Neeli

R. Prasad · Ramjee Prasad

Received: date / Accepted: date

Abstract In this paper, we propose a framework for performing Allocation,Assignment and Admission control (AAA) in next generation cellular net-works. A novel heuristic method for resource allocation is proposed. The al-location is done in a semi-distributed manner consisting of central allocation(CA) and local allocation (LA). The role of the assignment module is to esti-mate the amount of resources needed by a user in order to satisfy the qualityof service (QoS) requirements of the application. To that end, a Markov basedapproach which calculates the dropping probability of packets by consider-ing the effects of queuing in the medium access control (MAC) layer and theadaptive modulation and coding (AMC) in the physical layer is presented. Inorder to estimate the required resources, the predicted throughput and delayare calculated based on the dropping probability and the predicted values aremapped to the required ones. The admission control module is responsible foradmitting or rejecting a new or handoff user and is based on a mean resourcecalculation. The calculation takes into account the mean number of resourcesused by existing users as well as the buffer conditions of the individual users.By combining the three novel contributions on allocation, assignment and ad-mission control into the AAA framework the overall network as well as the

M. V. RamkumarAalborg University, CTIF, DenmarkTel.: +4599408606E-mail: [email protected]

Rasmus Hjorth NielsenCisco Systems, Denmark

Andrei Lucian StefanAalborg University, CTIF, Denmark

Neeli R. PrasadAalborg University, CTIF, Denmark

Ramjee PrasadAalborg University, CTIF, Denmark

2 M. V. Ramkumar et al.

cell-edge throughput have been improved and the number of admitted usershave been increased while still guaranteeing QoS for new users as well as ex-isting users.

Keywords Radio Resource Management · Admission Control · NextGeneration Cellular Networks · LTE

1 Introduction

Next generation cellular networks are facing severe challenges due to the ex-ponential increase in the number of mobile terminals and the variety of ap-plications [1]. One of the most challenging tasks in such networks is to meetthe quality of service (QoS) requirements of the users while improving the ef-ficiency of the overall network. This has to be achieved in the highly dynamicenvironments characterizing wireless networks which include rapidly changingfading, user mobility, traffic conditions, network conditions, etc. Future wire-less networks must be able to provide services to a vast number of mobile usersin different scenarios including both indoor and outdoor. In order to do so, cellsizes must be decreased which in turn results in an increase in the number ofbase stations needed to service a certain area [16]. At the same time, spectrumscarcity requires the reuse of resources among multiple cells and base stationsin both the up and downlink. This potential overlap of resource usage causesinter-cell interference (ICI) between cells, which decreases the QoS and alsothe cell throughput [3]. This problem is especially significant for cell-edge userswhich require high transmit power from the serving cell and thus potentiallycreating high interference on the neighboring cells utilizing the same resources.In order to address the aforementioned challenges, there is a strong need forefficient radio resource management (RRM).

RRM includes, among other tasks, the allocation of the right resources tothe individual cells in the network and in turn to the individual users served ineach cell [4] [5]. The allocation of resources should be done in such a way thatthe overall throughput of the network is maximized while still maintainingthe QoS of the users. If the resources are allocated orthogonally without anyoverlap between cells, the ICI will be zero, but the overall throughput is notmaximized due to lack of resources. On the other hand, a 1:1 reuse resultsin a high ICI which decreases the throughput. Therefore, a tradeoff on thepercentage of overlap should be obtained. As mentioned, users near the celledge are sensitive to interference as they are close to the neighboring cell andexperience weak signal strength due to path loss. Cell-center users, however,are not as sensitive to interference and the allocation of resources to each cellshould take these effects into consideration.

If the allocation is done in a decentralized and distributed manner, theneach cell does not have sufficient knowledge regarding the neighboring cellsand the allocation performed may not be optimal [6] [7]. The advantage of acentralized allocation is that it has knowledge of the overall network including

A Joint AAA framework for NGN 3

the load and user distribution in each cell [5]. However, the centralized alloca-tion has disadvantages in terms of signaling overhead [7]. A hybrid approachcould therefore provide an optimized solution and a tradeoff between the sig-naling overhead and the optimal allocation [5]. The amount of resources to beallocated to each cell and its users depends on the load of the cell and shouldbe determined prior to the actual allocation. The resources are a function ofthe QoS requirements, fading conditions, etc. of the users in the cell. Withrespect to maintaining the QoS of users, another important challenge is theadmission and allocation of resources to new users. New or handoff users candecrease the QoS for existing users by creating congestion in the network [1]and the admission should therefore be carefully considered.

The rest of the paper is organized as follows. Section 2 discusses the relatedwork and Section 3 explains the AAA framework and the interfaces betweenthe individual modules. Sections 4, 5, 6 explain the allocation, assignment andadmission control modules of the AAA framework. Section 7 is concerned withthe priority based scheduler (PBS) and Section 8 presents the simulation setupand the results obtained, while Section 9 concludes the paper.

2 Related Work

The allocation of resources can be done in a static way using fixed reuse meth-ods or in a dynamic way by considering the network conditions. Dynamicchannel allocation methods are less efficient than fixed allocation methods un-der high load conditions but provide more flexibility and traffic adaptability[7]. In fixed frequency reuse (FFR), physical resource blocks (PRBs) are allo-cated without overlap and hence the ICI is significantly lower, but at the costof reduced spectral efficiency. In [8] [9] [10], users in a cell are classified intodifferent classes based on their geometrical position and different bandwidthallocation patterns are assigned for different user classes. The most promisingapproach divides the users into two groups: interior cell-center users and exte-rior cell-edge users. One third of the available bandwidth in each cell is fixedfor cell-edge users and the rest for cell-center users [8] [9]. This approach iscalled soft frequency reuse (SFR). However, when the traffic load changes, itis desirable for the allocation of sub bands to cell-edge users not to be donestatically, but rather dynamically in order to take advantage of varying trafficload in the network. This is not addressed in [8] and [10]. In [9], an adaptiveSFR scheme dynamically adapts to changing traffic load and user distributionsamong neighbor cells. In [6], two methods for flexible spectrum usage (FSU)are proposed; spectrum load balancing (SLB) and resource chunk selection(RCS). In [11], partial frequency reuse (PFR) based on network load is pro-posed. In [5], a hybrid method for dynamic resource allocation is proposed. Inall the above mentioned methods, the allocation in the current frame is donebased on the signal to interference plus noise ratio (SINR) measurements ofthe previous frame. This approach does not guarantee the optimal allocationof resources for the current frame.


Whether the assignment or the admission modules are being discussed,the QoS requirements for a new user should be guaranteed without violatingthe QoS of existing users. The QoS achieved by a user depends on a seriesof factors, out of which the most notable are the channel conditions of theuser. The channel conditions cause packet errors and buffer overflow and thuspacket drops or increased delay in packet delivery [1]. Both channel and bufferconditions affect the throughput and delay of a user and traditional queuingmodels do not consider their effects. At the same time, channel models donot consider the effects of the status of the queue [12]. As an example, if thechannel is in deep fade, adaptive modulation and coding (AMC) will selecta lower modulation order, which will reduce the outflow of packets from thebuffer and thus the throughput of the user will be reduced. On the other hand,as the number of packets going out of the buffer reduces, the dropping rateof the packets increases which, in turn, will increase the delay. The previousexample was meant to show how the QoS experienced by the users is dependenton the channel and the queue characteristics.

The novelty of this work is the proposal of a joint AAA framework andthe system model explaining the interfaces between the modules. For centralallocation, a heuristic method of allocation is proposed, which works with thelocal allocation module in a semi-distributed way. A novel admission controlscheme based on mean resource method is proposed by taking into account thebuffer conditions of the users. The AAA framework and the proposed methodsare validated on a long term evolution (LTE) platform.

3 AAA FRAMEWORK

In this paper, a new framework for allocation, assignment and admission con-trol (AAA) is proposed as shown in Fig. 1. The main objectives of the proposedframework are:

– Dynamic and autonomous allocation of resources to each base station andto each user in the downlink by considering ICI, SINR, load of the network,location and downlink transmit power etc. such that the overall networkthroughput is maximized.

– Assignment of resources to each user, which includes estimating the numberof resources based on QoS requirements of the user, type of user, targetSINR, fading conditions, etc. such that the QoS requirements of the userare met.

– Admission of a new or handoff user such that the network does not experi-ence congestion, QoS of existing users is not violated and QoS of the newuser is achieved.

The allocation of resources in the network is done at two time scales, super-frame and frame. The central allocation (CA) is done for every super-frame bypredicting the SINR of the next super-frame from path loss and shadowing.Based on the predicted SINR, the CA allocates resources to each cell and


Fig. 1 AAA Framework

to each user for the next super frame. Accordingly, the CA module receivesinputs from the local allocation (LA) module. These inputs are the transmittedpower, the path loss and the amount of resources to be allocated to each usercalculated by the assignment module. The SINR of the next super-frame ispredicted from these inputs and, based on the predicted SINR, the CA isperformed.

In order to reduce the signaling overhead and to reduce the complexity,the CA is performed only once in a super frame. The goal of the CA is toimprove the overall network throughput by reducing the interference and toimprove the cell-edge throughput by providing lower reuse factors for cell-edgeusers compared to cell-center users. Due to the availability of data from allthe cells in the network the decisions taken by the CA are more effective froman overall network point of view.

The LA allocates the resources to each user in the particular cell based onthe channel and traffic conditions of the users. This operation is performed inevery frame. The CA recommends the resources to be allocated for each user,but the LA may change this allocation based on fading or buffer conditions ofthe user. The inputs received by the LA module from the other modules are:

– The users to be scheduled in the next frame (provided by the PBS).– The number of resources for every user (provided by the assignment mod-

ule).– Recommended resources to be used (provided by the CA).

The LA allocates resources to the users based on the SINR reported by theuser such that the overall throughput of the cell is maximized. The SINRvalues used at the LA level are the measured SINR values experienced by theusers in the previous frame.


The next module in the AAA framework is the assignment module whichis based on a two-dimensional Markov modeling of the queue. This model-ing takes into account the effects of AMC in the physical layer. Given that anew user is requesting admission to the network or that a handover has beeninitiated, the assignment module is triggered by the admission control (AC)module. The AC forwards the request on behalf of the user with the QoS re-quirements and channel conditions to the assignment module. The assignmentmodule then estimates the number of resources required by the user such thatits QoS requirements are met and it keeps providing this information to theCA module once every super frame and to the LA module every frame.

The next module of the AAA is the admission control module, which dealswith admission/rejection of a new or handoff user. For each new request, theassignment module calculates the number of resources required by the user andforwards it to the admission control module. The admission control moduleestimates the mean number of resources used by existing users by taking ad-vantage of multi user diversity based on buffer conditions. Based on the meannumber of resources used by existing users and the number of resources esti-mated for a new user (information provided by the assignment module), thismodule decides if admission is possible. The method of mean resource calcu-lation therefore increases the number of admitted users in the system withoutviolating QoS of existing users and hence decreases the dropping probability.

Another important module of the system model in Fig 1 that interfaceswith AAA framework is the PBS. Every frame, the PBS selects the usersto be scheduled based on the assigned priority and accordingly forwards thisinformation to the assignment module. The priority of the user is inverse pro-portional with the level of user satisfaction which is quantified by the achievedQoS level. The users with the highest priority are scheduled first in the nextframe. By scheduling the users with least satisfaction, fairness is obtained.

The goal of this section was to propose a generic framework (AAA) whichcan be used for any radio access technology (RAT) or for heterogeneous RATs.The modules and the interfaces between them have been described.

4 ALLOCATION

4.1 Central Allocation

In previous works [3] [4] [5], dynamic resource allocation is done based on theSINR experienced by the user in a previous frame or slot. This approach doesnot guarantee maximum throughput as the SINR changes completely for oneframe to another and there is not a function mapping the current SINR tothe previous SINR value or the previous allocations. Even if the allocationconverges or stabilizes after a few frames, this cannot be guaranteed to beoptimal as path loss and fading change due to user mobility, leading to newvalues of interference.


Hence a new allocation scheme is proposed, in which interference is pre-dicted by the central entity located inside the radio access network (RAN)[13] based on the current allocation of resources for the users. The allocationmodule receives input from the assignment module with a list of users anddownlink transmit power for each user, number of resources that need to beallocated in order to achieve the QoS and the path loss for each user. Theoutput of the allocation module is a mapping of available resources for eachbase station and the recommended resources for each user. Given that theuser has sufficient packets to be sent in his buffer or given that the user isnot experiencing deep fade, the base station may follow the recommendationof the CA module. The base station also has the possibility of taking its owndecision about the users to be scheduled and the resources to be allocated ateach frame level. This local decision depends on traffic conditions of the user,the channel fading conditions and the user satisfaction levels.

Furthermore, the interference experienced by a user in the next frame isestimated from the channel gain which is based on the path loss of the userfrom its own base station and neighboring base station. The allocation ofresources first considers the users experiencing the highest downlink transmitpower (which are most probably cell-edge users) and thus potentially creatingthe most interference in the system. Once the resources for the user have beenallocated, the effect of this allocation on other users in the system is calculated.

The reason for allocating the far off users first is that the interferencecreated by far off users is higher compared to the nearby users. This way ofallocating resources, based on geographic position, ensures a low reuse factor(e.g. 1/3) for cell-edge users and a high reuse factor (e.g. 1) for cell-center users.The increase in reuse factor from cell-edge to cell-center increases dependingon the load of the network and distribution of the load in the network. Theproposed heuristic method of CA is explained for an orthogonal frequencydivision multiple access (OFDMA) scheme based system (LTE).

Consider a network consisting of L base stations and a set of M users, withusers served by base station l denoted as Ml, where M =

⋃L

l=1 Ml. The basestation l ∈ L that serves user m ∈ M is denoted as l(m). Let Gm,j denotelong term channel gain, which includes path loss and shadowing, from basestation j ∈ L to user m ∈ M . Gm,j is calculated from dm,j as shown in eq. 1,where dm,j is the distance from user base station j to user m. The downlinktransmit power of user m from l(m) is denoted as Pm.

Gm,j =(

44.9−6.55log10(hBS))

log10(dm,j)+34.46+5.83log(hBS)+23(fc/5)(1)

The multiple access scheme used is OFDMA with K PRBs in each frame,with user m allocated with a set of km PRBs, of length |km| obtained fromthe assignment module, as shown in Fig. 2 where

∑

m∈M

|km| ≤ K (2)


Fig. 2 Physical resource blocks of OFDMA system

The goal of the CA module is to find km for all m ∈ M .The method for CA is as follows. All users in the network, M , are arranged

in decreasing order of their downlink transmit power Pm and they are allocatedin the same order. Let i be the user to be allocated. The interference PIim dueto user i on user m in the system is predicted as

PIim = Pi.Gm,j (3)

where m ∈ M,m 6= iHence the total interference Ii seen by user i, from all users in Y that were

already allocated is calculated as in eq. 4. In eq. 4, km is the PRBs allocatedto user m ∈ Y and PImi (km) is the interference seen by user i on km PRBsalready allocated to user m in the list. Initially Ii = 0 on all the PRBs andwhen the users are allocated, Ii is updated by adding the interference fromthe already allocated users on their corresponding PRBs. Until all the PRBsare allocated once, each user gets unused PRBs and thus interference is non-existent. Once all of the PRBs are allocated, there is a need for reusing theresources in which case the next users in the list face interference from theusers that were already allocated with the same PRBs and viceversa.

Ii =∑

m∈Y

PImi (km) (4)

Using eq. 4 and eq. 5, the SINR of user i can be calculated by taking intoaccount the transmit power allocated to the user, the path loss experiencedand the sum of the interferences and the corresponding noise (eq. 5).

SINRCAi =

Pi.Gi,j

Noise+ Ii(5)

where i /∈ Y, j = l(i) Once SINRCAi and PIim are calculated for user i, a

ratio Ri is defined on each PRB which is the SINRCAi of user i on each PRB

divided by the total interference exerted by i on other users in the system oneach PRB:

Ri =SINRCA

i∑

m∈M,m 6=i PIim(6)

where i /∈ Y


The PRBs to be allocated for user i, ki, are selected such that the ratioRi is maximized. This ensures that the user is allocated the optimum PRBswith a good SINR and the total interference exerted by user i on the systemis minimum. Once ki is found the interference due to user i on other users inM on ki PRBs is updated, and Y is updated with i. For the next user in thelist the same procedure is repeated, until all the users are allocated.

This method reuses the PRBs if the load in each cell increases, such that theinterference from the users already allocated is minimum and the interferencecreated by this user on other users in the system is also minimum. Hencethis method guarantees that, when PRBs are reused, the users in the cell-center obtain a lower reuse factor and the users near the cell edge experience ahigher reuse factor. The proposed method can be applied to cells with unevendistribution of loads. This heuristic approach improves the overall throughputof the network and guarantees cell-edge throughput, which is verified fromsimulation results shown in Section 8.2. This approach is suitable in a networkwith dynamic variation of traffic conditions and network load.

The complexity of the method in terms of the number of multiplications isconsidered. For theM th user, eq. 4 needsM−1 multiplications. By consideringdivision also as multiplication, eq. 5 needsK multiplications, one for each PRB.Similarly denominator of eq. 6 needs M − 1 multiplications and the divisionrequires K multiplications. Hence the total number of multiplications for theM th user is 2(M + K − 1). Hence the complexity of the proposed heuristicmethod grows linearly with M and K.

4.2 Local Allocation

Due to the signaling overhead, the CA module takes the path loss and shad-owing into consideration but not the fading effects of users. Hence this mod-ule takes advantage of channel conditions by allocating resources having highSINR to users. Even though the CA module in the RAN recommends theresources to be used by each user in every super frame, the LA may changethe allocation by considering the changes in signal conditions. This moduletakes the input from the PBS and from the assignment module and allocatesresources to each user. The allocation of resources to the users is based on theSINRLA

i,k reported to the LA module of user i on PRB k, which also takeschannel gain due to fading into consideration. Hence the goal of this moduleis to maximize the overall SINRLA

i,k of the base station by allocating resources

to the users having the best SINRLAi,k conditions, which in turn maximizes the

overall throughput of the base station.

eq. (7) explains the SINR measured by the user where hki,l(i) is the channel

gain due to fading between user i and its own base station l(i) on PRB k.|Ml| is the total number of users to be scheduled in base station l with useri allocated with |ki| PRBs and P k

l is the downlink transmit power from base


station l on PRB k.

SINRLAi,k =

P kl(i).Gi,j .h

ki,j

Noise+∑

l∈L,l 6=l(i) Pkl .Gi,l

(7)

The goal of the LA is to find ki PRBs for user i such that the overall SINRLAi,k

in each base station is maximized, as shown in eq. (8).

max

(

|Ml|∑

i=1

ki∑

k=1

SINRLAi,k

)

where ki ∈ K, |ki| < |K|,Ml ∈ M, |Ml| < |M |

(8)

5 Assignment

The assignment module estimates the number of resources required by eachuser in order to have its QoS requirements met. This estimation is basedon a Markov-based modeling of the queue in the MAC layer by taking intoaccount the effects of AMC in the physical layer. Hence, this module guaranteesmeeting the QoS requirements for the existing users and also for the new useras this module is triggered at the time of admission. The assignment modulegets input from the PBS regarding the users to be scheduled in the next frameand it forwards to the LA the number of resources required for every user.Every super-frame it also sends to the CA module the user information, thenumber of resources required by each user, downlink transmit power of theuser and the path loss of the user.

The estimation of the amount of resources takes into account the AMC inthe physical layer and the effects of the queue in the MAC layer. Each useris given b resources in each frame and the goal of the assignment module isto find suitable value of b that guarantees the QoS to the user. By using aMarkov-based analysis for the queue [16] as shown in Fig. 3, the assignmentmodule estimates the probability of dropping a packet (Pd). From the droppingprobability the estimation for throughput and delay for different values of bare then derived. The most suitable value of b that matches the requested QoSis selected.

Each state of a Markov chain is defined as (U,C) where U is the number ofpackets waiting in the queue (can range from 0 to B where B is the buffer size)and C represents the number of packets transmitted in the next frame. Thenumber of packets transmitted in a frame depends on the number of resourcesallotted to the user and the AMC mode of the user in that particular frame.

Cn = bBn n = 1, 2...N (9)

where Bn is the number of bits per symbol depending on the AMC mode.Hence C can take any value in C1, C2, ..., CN where N is the number of AMCmodes.


Fig. 3 State transitions of two-dimensional Markov chain of a queue

Each user is allotted a buffer in the MAC layer and the size of the bufferB depends on the type of user. For high data rate applications like videoconferencing, the buffer size B is large compared to low data rate applicationslike voice. The number of packets waiting in the queue depends on the arrivalprocess A, the service process and the buffer length allocated to the user. Thearrival process is modeled with a Poisson distribution:

P (a) =λae−λ

a!(10)

where a ≥ 0 and E(A) = λ is the packet arrival rate, defined as the averagenumber of packets arriving during one frame, depending on the traffic model.

The service process depends on the AMC and on the SINR of the userin the frame. The probability of the service process changing from one stateto another depends on the transition probability of a user changing from oneAMC mode to another by assuming that the number of resources allottedto the user is fixed. In [14], the signal to noise ratio (SNR) was divided intoadjacent regions based on the desired bit error rate (BER). The transitionprobabilities between the various SNR regions were determined based on theLevel Crossing Rate (LCR) of the channel fading distribution. Also it is as-sumed that a user can transition one AMC mode at a time which can be seenin the two-dimensional Markov chain of the queue in Fig. 3

Based on the probability of packet arrival, P (a), from the arrival processand the transition probability between AMC modes, the steady state distribu-tion of the two-dimensional Markov chain is calculated. From the steady statedistribution of the two-dimensional Markov chain P (U = u,C = c) [15], theexpected number of packets dropped from queue E(D) can be expressed as

E(D) =∑

a∈A,u∈U,c∈C

max(0, a−B+max(0, u−c))P (A = a)×P (U = u,C = c)

(11)


The dropping probability of packets, Pd, from the queue is calculated fromthe expected number of packets dropped from the queue and expected arrivalrate of packets as

Pd =E(D)

λTf

(12)

where Tf is the frame duration. From the above dropping probability, Pd, thepacket loss rate (PLR) is calculated, which is defined as the probability that apacket is lost and is composed of two factors: packet error rate (PER), whichis the ratio of packets lost due to radio environment, noise, etc. and the packetdropping rate (PDR), which is the ratio of packets dropped due to timeoutsin the queue or due to the finite buffer length B. The packet is assumed to belost when there are bit errors in the packet and/or when the waiting time of apacket in the buffer is more than a certain timeout. The PLR is calculated as

PLR = 1− (1− Pd)(1− P0) (13)

where P0 is the PER due to channel fading and Pd is the dropping probability.From PLR, the prior throughput is estimated as:

ηprior = λ(1 − PLR) (14)

From Little’s Theorem [17], the average number of packets waiting in the queueis equal to the product of arrival rate of the packets and the average delay ofeach packet. From this the expected prior delay is estimated as

τprior =Nw

E(A)(1 − Pd)(15)

where Nw is the average number of packets waiting in the queue plus theaverage number of packets transmitted in one frame obtained from the steadystate probability P (U = u,C = c). ηprior and τprior are calculated for differentvalues of b. The minimum value of b that achieves the required QoS in termsof throughput and delay requested by the user is selected. The assignmentmodule is triggered during admission to a new user and hence the estimatedvalue of b is sent to the admission control module. It is assumed that delayin the transmission is only due to waiting time in the buffer, hence only Pd isconsidered for the delay calculations. The delay caused due to retransmissionscaused by CRC errors is not considered.

6 Admission Control

The admission control [1] module is triggered when an admission request froma new user is received. The user sends a request with the QoS requirementsneeded for its application such as target SINR, data rate, delay, PER andchannel conditions (fading rate fd and fading index m). After receiving theseinputs the assignment module estimates the amount of resources needed toobtain the QoS requested. Thus the assignment module checks whether the


required resources are available in the network and accordingly takes the de-cision to either admit or reject the user.

The admission control algorithm increases the number of connections thatcan be served by taking the buffer conditions of each user into consideration[12]. A user may not need to utilize all the resources allocated due to lackof packets in the buffer. Thus, in order to obtain an efficient utilization ofthe bandwidth, a mean resource calculation which finds the average numberof resources used by all the users in the system is performed. The numberof resources actually scheduled [11] can be expressed in the following way,depending on the current channel conditions and the previous buffer status:

k(Ut−1, Ct) =

0; if Ct = 0,

km; if Ut−1 ≥ Ct,

f loor(km∗Ut−1

Ct

); if Ut−1 < Ct.

(16)

where Ut−1 is the number of packets that are in the queue for user m at thetime moment t−1, km is the number of resources estimated by the assignmentmodule for user m and Ct is the number of packets that can be accommodatedin the next frame with the selected AMC mode. If km is the maximum numberof resources that can be allocated to user m, then the maximum number ofresources that can be allocated to all the users in a system is kM =

∑

m∈M km.The users are admitted until kM reaches the maximum number of resourcesin the system. The goal of the mean resource allocation is to find the averagevalue of kM , such that a maximum number of users can be accommodated inthe system.

The mean number of resources used by all users, or average value of kM , isestimated from the steady state distribution of kM , which can be determinedfrom the Z-transform Dm(z) of km. The Z-transform of km can be expressedas:

Dm(z) =∑

j

P (km = j)z−j (17)

where P (km = j) is the probability that the user m is allocated with j re-sources. The Z transform of kM is expressed as DM (z) =

∏

m∈M Dm(z). Bycalculating the inverse Z-transform, the steady state distribution of kM is ob-tained as:

P (kM = j) = Z−1{

DM (z)}

(18)

From the steady state distribution of kM the mean number of resources kmeanM

used by all the users in the system is obtained. Based on the estimated valueof kmean

M , a new user m which requires km resources is admitted according to:

km + kmeanM ≤ Ktotal (19)

where Ktotal is the total number of resources available in the system.


Table 1 Weight Coefficients

Service type Notes ωrt

uωnrt

u

1 High rate and low delay 1 02 Low delay 1 13 High rate 0 14 Best Effort 0 0

7 Priority Based Scheduler

The main function of the PBS is to schedule the users based on the estimatedpriority so that the user with highest priority is scheduled first. The amountof resources to be allocated to each user is estimated by the admission controlalgorithm based on the achieved QoS. For each user a satisfaction index (SI)which gives the level of user satisfaction and indicates how throughput anddelay of a user is achieved w.r.t. the desired values is calculated. The desiredQoS values are assumed to be dependent on the type of service. From the SIvalues the PBS calculates the priorities for each user and sends them to theLA module. The SI is represented as a function of delay Γu(t) or as a functionof rate Ψu(t) [12]. In either case, the lower the SI, the higher the priority userwill be assigned.

The delay component SI is expressed with regards to the head of line (HOL)delay ωu, which is the longest delay experienced by a packet at the HOL, andthe maximum delay for service u, T (u), as shown below:

Γu(t) =

{

T (u)−∆T (u)ωu(t)

; if ωu(t) < T (u)−∆T (u),

1; otherwise.(20)

where ∆T (u) is a safety margin. The rate component is expressed in terms ofthe average rate measured, ηu, and the desired data rate, η̂u:

Ψu(t) =ηu

η̂u −∆η̂u(21)

where ∆η̂u is the margin coefficient. The safety margin and margin coefficientare used due to the variations in the radio link conditions.

The priority function has two components Φrtu and Φnrt

u as shown in eq. 23and eq. 24, where Φrt

u is the real time component based on the delay and Φnrtu

is the non real time component based on the data rate.

Φu = ωrtu φrt

u + ωnrtu φnrt

u (22)

where the weight coefficients ωrtu and ωnrt

u are determined based on the servicetype as shown in Table 1. The expressions for Φrt

u and Φnrtu are

Φrtu =

Ru1

Γu(t); if Γu(t) ≥ 1;Ru 6= 0,

1; if 0 < Γu(t) < 1;Ru 6= 0,

0; if Ru = 0

(23)


Φnrtu =

Ru1

Ψu(t); if Ψu(t) ≥ 1;Ru 6= 0,

1; if 0 < Ψu(t) < 1;Ru 6= 0,

0; if Ru = 0

(24)

where Ru is the normalized channel quality in the range [0 1], as high receivedSNR induces high capacity which results in high priority.

Using the above equations, the calculation of the priority function Φu isperformed. From the above calculations, the users with least satisfaction aregiven the highest priority and scheduled first. This approach ensures fairnessfor each user.

8 Simulation Results

8.1 Simulation Setup

The setup simulated in MATLAB is a 4-cell network, with users distributedrandomly in each cell as shown in Fig. 4. LTE-TDD is used with 10MHzbandwidth at 3.5GHz center frequency. The number of PRBs used in each celldepends on the load of the cell. The PRBs are distributed uniformly amongusers in the cell based on the type of the user. The users are moving with 3km/h speed within the cell with a 250m radius in urban scenario. The usersmove straight in random directions and bounce back when a cell boundaryis reached. The path loss model used for the urban scenario is C2 NLOSdeveloped in Winner [18] as shown in eq. 1.

Each user is assigned a buffer of length 15, 30 or 60 packets dependingon the type of user. Three types of users are selected based on the data rate400kbps, 800kbps or 2Mbps. Based on the type of user the λ of the poissondistribution is chosen, which decides the packet arrival rate. The number ofusers depends on the load of the cell with each type of user randomly selected.

The transmit power of the base station is 43dBm, which is equally dis-tributed on all PRBs. In each frame the SINR is measured on each PRB bycalculating the received signal from the current base station and interferingsignal from neighbor base stations on each PRB. Based on the measured SINRthe LA assigns the PRBs to each user such that the overall SINR in each cellis maximized.

The CA module allocates the PRBs to each cell once in every super framebased on the predicted SINR. A super frame consists of 20 frames. The pro-posed CA+LA is compared with an existing method [5] in the literature whichis also a hybrid approach, with a standalone LA complemented by a randomallocation. The base station transmits to each user on the allocated PRBs. TheSINR is measured in each frame and mapped to the Shannon throughput. Thesimulation parameters used in the simulation are given in the Table 2. Thesimulation is run for 1000 frames and averaged over 100 runs.


Fig. 4 4-cell network

Table 2 Simulation Parameters

Parameter Value

Carrier frequency fc 3.5GhzDeployment scenario Urban macroIntersite distance 500mPath loss model C2 NLOS

Mobility 3 KmphBandwidth 10 MhzAntenna Omni directional

downlink Tx power 43 dBmUL Tx power 24 dBmtarget SINR 15 dBNoise figure 9dB

Buffer length B [15, 30, 60] for Type 1,2,3Packet length 84 bits

Packet arrival rate, λ [3, 5, 12] for Type 1,2,3Number of AMC modes BPSK, 4, 8, 16, 32, 64 QAMSINR thresholds [dB] [7.2 10.1 12.5 16.1 18.7 22.2]

8.2 Cell Throughput

Fig. 5.a shows the mean cell throughput and average cell-edge throughput ofthe network with increasing load. The load is varied by varying the numberof PRBs used in the cell. The existing hybrid method is compared with theproposed CA+LA method, and it can be seen that at 70% and 80% load, theproposed approach performs around 2Mbps or 14% better than the existingmethod.


Also with CA+LA the performance of mean cell throughput is around4Mbps better than LA alone at 60% and 70% loads. The CA recommends thePRBs to be used by each cell for every super frame and, as load reaches 100%,the CA has to recommend all of the PRBs for each cell; hence the performanceof CA+LA and LA standalone merges at 100% load. At lower loads, as theeffect of interference is low, the difference in performance between CA+LA andLA standalone is lower when compared to higher loads. The proposed solutionis also compared with a random allocation. It can be seen that LA standaloneand CA+LA perform around 5Mbps better than random allocation at 100%load. At 90% load CA+LA performs 6Mbps better than random allocation,whereas LA standalone performs 4Mbps better than random allocation.

The cell-edge throughput shown in Fig. 5.b is defined as the outage through-put of users below 5%. From the CDF of the average user throughput of allusers, the value at 5% is outage throughput, which is assumed as cell edgethroughput. The performance of cell-edge user for the proposed approach isbetter by 125kbps or around 30% compared to the existing method [5] at 60%load. As explained above, the difference in performance can be better observedat medium loads. At 100% load the performance of the proposed heuristicmethod for CA with LA and the LA standalone is the same. At 90% load, theproposed CA+LA performs around 400kbps better than random allocation,whereas LA standalone performs 300kbps better than random allocation. Thedip at lower loads is due to less or zero number of cell-edge users at the lowerloads. Due to the less number of users, cell edge throughput is low at lowerloads.

8.3 User Throughput

The assignment module estimates the number of PRBs required by each typeof user for a given target SINR. We assume three types of users with datarates 400kbps, 800kbps and 2Mbps. For each user type the number of PRBsrequired in order to obtain the target throughput is estimated by the assign-ment module. For each user type the parameters chosen are shown in Table 2.For AMC, if the SINR is below 7.2dB, then there will be no transmission.

By using the above values the assignment module estimates the probabilityof dropping the packets for various values of b, where b is the number ofPRBs allotted to each user. From the dropping probability the throughput iscalculated. Fig. 6.a, Fig. 6.b and Fig. 6.c show the average throughput of type-1, type-2 and type-3 users in the system with required throughput of 400kbps,800kbps and 2Mbps respectively. The red line shows the required throughputof the user and the blue line shows the achieved throughput. It can be seen thaton an average all users in each type obtain the required throughput. Hencethe assignment module guarantees the QoS for the user.


20 30 40 50 60 70 80 90 1004

6

8

10

12

14

16

18

Load in %

Th

rou

gh

pu

t in

Mb

ps

Proposed CA+LALA aloneExisting methodRandom allocation

20 30 40 50 60 70 80 90 1000

100

200

300

400

500

600

Load in %

Th

rou

gh

pu

t in

Kb

ps

Proposed CA+LALA aloneExisting methodRandom allocation

Fig. 5 Comparison of a) mean cell throughput and b) cell-edge throughput

8.4 Mean Resource Evaluation

Fig. 7 and Fig. 8 show the performance of admission control with mean re-source algorithm. Fig. 7 shows the average user throughput for the type-2and type-3 users with an increasing number of users, with and without meanresource calculation. Without mean resource calculations, the maximum num-ber of type-3 users that can be admitted in the network is 40 (10 users foreach cell). The red curve illustrates the average user throughput with meanresource calculations and it can be seen that 44 users are admitted in thesystem, giving a 10% increase in number of users compared to the scenario


0 10 20 30 40 50 60 70 80 90 100

400

450

500

550

Number of frames

Ave

rag

e U

ser

Th

rou

gh

pu

t in

Kb

ps

Average user throughput for type−1 userMinimum throughput requirement for type−1 user

0 10 20 30 40 50 60 70 80 90 1001.8

2

2.2

2.4

2.6

2.8

Number of frames

Ave

rag

e u

ser

thro

ug

hp

ut in

Mb

ps

Average user throughput for type−3 userMinimum required throughput for type−3 user

Fig. 6 Average user throughput of a) type-1 b) type-2 and c) type-3 users

in which no mean resource calculation is performed. It can be seen that withmean resource calculation the number of users admitted in the system in-creases while maintaining the average user throughput. Hence, the admissioncontrol algorithm guarantees QoS for the new user while maintaining QoS forexisting users, which can be seen in Fig. 7 as the number of users increases asthe load increase from 0 to 100%.


0 20 40 60 80 100 1207.5

8

8.5

9

9.5

10

10.5

11x 10

5

Nr of users

Th

rou

gh

pu

t in

bp

s

Throughput with mean resource calculationThroughput without mean resource calculationThroughput requirement for type 2 users

0 5 10 15 20 25 30 35 40 45 50

2

2.1

2.2

2.3

2.4

2.5

2.6

2.7

2.8x 10

6

Nr of users

Th

rou

gh

pu

t in

bp

s

Throughput without mean resource calculationThroughput with mean resource calculationThroughput requirement for type 3 users

Fig. 7 Mean resource performance for a) Type-2 and b) Type-3 users

In case of type-2 users, it can be seen that with mean resource calculationthe number of users admitted in the system is increased by 10% while theaverage user throughput is maintained and the minimum QoS requirement ismet. In both cases, the average user throughput decreases marginally with anincrease in the number of users. This is due to the fact that more users have to


Fig. 8 Dropping probability vs Erlang load for Type-2 and Type-3 users

share the same resources. Nevertheless, the minimum required throughput isstill achieved even in the case in which more users are admitted to the system.

Fig. 8 illustrates the dropping probability with the variation of the ErlangLoad for type-3 and type-2 users. The Erlang load is defined as the productof the call arrival rate and mean duration of the call. It can be seen that byapplying the mean resource calculation, the dropping probability is reduced byaround 5.5% for type-3 user and 2.2% for type-2 user when compared to thescenario without mean resource calculation at 110% Erlang load. The increasein performance of mean resource calculation is highlighted better for type-3users compared to type-2. This is due to the fact that type-3 users require ahigher number of PRBs than type-2 users, which increases the probability ofunused PRBs for type-3 users compared to type-2 users. Hence the meanresource algorithm increases the number of admitted users, which in turnreduces the dropping probability. For loads below 80% the difference in theperformance between the curves is insignificant due to lack of averaging.

The mean resource calculation for type-1 users is not performed, as thelength of the output of two finite sequences of length l1 and l2 is l1 + l2− 1.Hence if l1 + l2 − 1 is greater than one then the mean value of the area canbe calculated, as the number of users are being admitted. For type-1 userthe Dm(z) is always a delta function, hence the output of convolving for anynumber of users is always a delta function, for which the mean value is alwayszero. Hence the mean resource calculation is not possible to perform whenthe system has type-1 users alone. Also the mean resource calculation is notperformed for mixture of different types of users, as the increase in the numberof users cannot be seen clearly for mixture of users.


Table 3 Overhead Comparison

Scenario Overhead in kbps

1 - less info every frame 9202 - less info every super frame 46

3 - full info every frame 64004 - full info every super frame 320

8.5 Overhead Analysis

The CA module receives inputs from different base stations in the network.The amount of overhead information sent by each base station is estimated forfour scenarios. In scenarios 1 and 2 less information which includes userids,number of resources, downlink transmit power, path loss, is sent once forevery frame and super frame respectively. In scenarios 3 and 4 full informationwhich includes the channel gain information on all the PRBs apart from theinformation in scenario 1 or 2, is sent once for every frame and super framerespectively. eq. 25 and eq. 26 show the overhead with less information forscenarios 1 and 2, OVless, and overhead with full information for scenarios 3and 4, OVfull respectively. Here it is assumed that 64 bit IMSI (Internationalmobile subscriber identity) is used as user id and 32 bits are used for floatingpoint representation of path loss and downlink transmit power.

OVless = 64 ∗ |M |+ floor(log2(K)) ∗ |M |+ 32 ∗ |M |+ 32 ∗ |M | ∗ |L|

(25)

OVfull = 64∗|M |+floor(log2(K))∗|M |+32∗|M |+32∗|M |∗|L|+32∗|M |∗K(26)

where |M | is the number of users, K is the number of PRBs and |L| is thenumber of base stations.

The overhead values calculated with the parameters used in the simulationfor four scenarios are shown in Table 3. It can be seen that scenario 2 has theleast overhead (46kbps). The overhead of scenario 3 is higher compared to theother three scenarios, which is impractical from an implementation point ofview. In scenario 2, although there is an overhead of 46kbps due to CA, theproposed method gives 4Mbps improvement in overall cell throughput whencompared to the LA standalone solution. Hence the overhead of 46 kbps canbe justified.

9 Conclusion

The proposed AAA framework considers the combination of the three mainaspects of RRM - allocation, assignment and admission control. In proposingthe joint framework, we have successfully managed to address many of thechallenges experienced by next generation cellular networks.


The proposed heuristic method for allocation of resources adapts to dif-ferent load conditions and distribution of users across multiple base stations.The prediction of SINR in the next super frame guarantees that the new allo-cation performed centrally (CA) is optimal and based on both path loss andshadowing while small scale fading effects are considered locally (LA). Theproposed system model considers the traffic, channel, buffer conditions andQoS requirements of all users in order to estimate the required resources. Byusing the Markov based estimation of resources required by each user, band-width utilization is improved in terms of efficiency. At the same time the QoSrequirements for both new as well as existing users are met. Furthermore,the proposed admission control method increases the total number of usersadmitted in to the system without violating the QoS of users.

Based on the simulation results from an LTE network, it can be con-cluded that the AAA framework provides better overall network and cell-edgethroughput than comparable methods. The framework also increases the to-tal number of users in the system, hence decreasing the dropping probabilityof new and handoff users while guaranteeing QoS for existing users. Fromthe overhead analysis it can be seen that the improved performance does notcome at the expense of increased overhead. The framework can be extended toany next generation wireless network and for heterogeneous network scenariosthrough common RRM approaches.

Acknowledgements The authors would like to thank ”Fibre-Optic Networks for Dis-tributed Extendible Heterogeneous Radio Architectures and Service Provisioning (FUTON)”- an EU funded FP7 project (ICT-2007-215533) and ”Research and Development on Con-verged network of wireless and wired systems using frequency sharing type wireless tech-nologies” - a research project funded by NICT, Japan.

References

1. Ramkumar. M. V, Bayu Anggorati, Andrei Lucian Stefan, Neeli R. Prasad, RamjeePrasad. QoS-Guaranteed admission control for OFDMA-based systems. Globecom 2010;606-610. DOI: 10.1109/GLOCOMW.2010.5700392

2. FP7 ICT Project FUTON, at http: www.ict-futon.eu.3. Elayoubi SE, Ben Haddada O, Fourestie B. Performance evaluation of frequency plan-

ning schemes in OFDMA-based networks. IEEE transactions on wireless communications2008; 7(5): 1623-1633. DOI: 10.1109/TWC.2008.060458.

4. Guoqing li, Hui Liu. Downlink radio resource allocation for multi-cell OFDMA sys-tem. IEEE transactions on wireless communications 2006; 5(12): 3451-3459. DOI:10.1109/TWC.2006.256968.

5. Koutsimanis C, Fodor G. Dynamic Resource Allocation Scheme for Guaranteed Bit RateServices in OFDMA Networks. International Conference of Communication WorkshopsICC 2009; 2524-2530. DOI: 10.1109/ICC.2008.478.

6. Kumar S, Yuanye Wang, Marchetti N, Kovacs IZ, Pedersen KI, Mogensen PE. SpectrumLoad balancing for Flexible spectrum usage in Local area deployment scenario. 3rd IEEESymposium on New Frontiers in Dynamic Spectrum Access Networks DySPAN 2008; 1-5.DOI: 10.1109/DYSPAN.2008.93.

7. Katzela I, Naghshineh M. Channel Assignment schemes for cellular mobile Telecommu-nication systems: A comprehensive survey. IEEE personal communications June 1996;3(3): 10-31. DOI: 10.1109/98.511762.


8. E3 project deliverable9. Xuehong Mao, Maaref A, Koon Hoo Teo. Adaptive Soft Frequency Reuse for Inter-Cell

Interference Coordination in SC-FDMA based 3GPP LTE Uplinks. Globecom 2008; 1-6.DOI: 10.1109/GLOCOM.2008.ECP.916.

10. R1-050507, ”Soft frequency Reuse scheme for UTRAN LTE”; Huawei, 3gpp TSG RANWG1 Meeting 41, Athens , Greece, May 2005.

11. Krasniqi B, Wrulich M, Mecklenbrauker CF. Network-load dependent Partial frequencyreuse for LTE. International Symposium on Commnucations and Information TechnologyISCIT 2009; 672-676. DOI: 10.1109/ISCIT.2009.5341160.

12. Qingwen Liu, Xin Wang, Giannakis GB. A Cross-Layer Scheduling algorithm with QoSSupport in Wireless Networks. IEEE Transactions on Vehicular technology 2006; 55(3):839-847. DOI: 10.1109/TVT.2006.873832.

13. IEEE 1900.4. Architectural building blocks enabling network-device distributed decisionmaking for optimized radio resource usage in heterogeneous wireless access networks

14. Q. Zhang, S. Kassam. Finite-state Markov model for Rayleigh fading channels. IEEETransactions on Communications 1999; 47(11): 1688-1692. DOI: 10.1109/26.803503.

15. Qingwen Liu, Shengli Zhou, Giannakis GB. Cross-Layer Scheduling With PrescribedQoS Guarantees in Adaptive Wireless Networks. IEEE Journal on selected areas 2005;23(5): 1056-1066. DOI: 10.1109/JSAC.2005.845430.

16. Qingwen Liu, Shengli Zhou and Georgios B. Giannakis. Queuing With Adaptive Mod-ulation and Coding Over Wireless Links: Cross-Layer Analysis and Design, IEEE Trans-action on Wireless Communications, 4(3):1142-1153, may 2005

17. R. B. Cooper, Introduction to Queuing Theory, 2nd ed. Elsevier North-Holland, 198118. IST-4-027756 WINNER II, D1.1.2, WINNER II Channel Models part I, September

2007.


AAA Allocation Assignment and Admission controlAC Admission controlAMC Adaptive Modulation and CodingBER Bit Error RateBS Base StationBW BandwidthFFR Fixed Frequency ReuseFSU Flexible spectrum usageHz HertzICI Inter Carrier Interference

KBps Kilo byte per secondkbps kilo bits per secondLTE Long Term EvolutionMAC Medium Access Controlmbps mega bit per secondms millisecond

OFDMA Orthogonal Frequency-Division Multiple AccessPBS Priority based schedulerPER Packet Error RatePHY Physical LayerPRB Physical resource blockPLR Packet Loss RateQAM Quadrature Amplitude ModulationQoS Quality of ServiceRAN Radio Access NetworkRAT Radio Access TechnologyRRM Radio Resource ManagementSFR Soft Frequency ReuseSNR Signal-to-Noise RatioSINR Signal to Interference plus Noise Ratio

AAA (Allocation, Admission and Assignment Control) for Network management

Education

Transcript of AAA (Allocation, Admission and Assignment Control) for Network management