A novel QoS aware medium access control scheduler for LTE-advanced network

Accepted Manuscript

A Novel QoS aware Medium Access Control Scheduler for LTE-Advanced Network Saptarshi Chaudhuri , Irfan B , Debabrata D PII: DOI: Reference:

S1389-1286(18)30035-5 10.1016/j.comnet.2018.01.024 COMPNW 6372

To appear in:

Computer Networks

Received date: Accepted date:

7 May 2017 17 January 2018

Please cite this article as: Saptarshi Chaudhuri , Irfan B , Debabrata D , A Novel QoS aware Medium Access Control Scheduler for LTE-Advanced Network, Computer Networks (2018), doi: 10.1016/j.comnet.2018.01.024

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

A Novel QoS aware Medium Access Control Scheduler for LTE-Advanced Network

CR IP T

Saptarshi Chaudhuri*, Irfan Baig*, and Debabrata Das** *Tata Power SED, Electronic City, Bangalore – 560100, India. **International Institute of Information Technology-Bangalore, Electronics City, Bangalore – 560 100, India *{schaudhuri, ibaig}@tatapowersed.com ; **[email protected] To realize the application packet flow within the eNodeB system, in Fig-1 we also show the data plane protocol architecture as mandated by [1]. The data plane architecture consists of the protocol layers like physical (PHY), medium access control (MAC), radio link control (RLC), Packet Data Convergence Protocol (PDCP), and GPRS tunneling protocol user plane (GTPU). According to [1], if N number of users open M applications towards the internet, then there would be N x M numbers of logical channels created in RLC layer, with N x M queues (each application mapped to one queue) created at the MAC layer and N number of GTPU tunnels created at eNodeB system. These GTPU tunneled data is exchanged over the S1 user plane interface as shown in Fig-1.

ED

M

AN US

Abstract—LTE-A specification mandates Medium Access Control (MAC) scheduler entity to ensure strict guaranteed quality of service (QoS) both in downlink and uplink direction. To the best of our knowledge the LTE-A MAC schedulers proposed so far constitute of an objective function dependent on single constraints like number of radio resources, user throughput, user channel conditions, user data buffer or user’s downlink power requirement. In real world scenario if all the above constraints are not simultaneously taken into consideration, then it would be difficult to meet LTE-A QoS requirements. In this regard, to the best found knowledge, for first time in this paper, we propose a Multi Objective QoS aware LTE-A Downlink-MAC Scheduler (MOQDS) algorithm which adhere to two level QoS and fairness requirements of LTE-A specification. MOQDS algorithm does scheduling at two levels with each level has its objective function with its multiple operational constraints. Every transmission time interval, MOQDS uses multi-objective optimization to selects right user(s) and its corresponding application(s) to meet LTE-A QoS requirements. Simulation results are being compared with well referred LTE-A schedulers like modified largest weighted delay first, exponential rule proportional fairness and log rule under various ITU channel models. MOQDS achieves an average of 50% reduction in packet drop rate and minimum three times increased cell throughput as compared to above mentioned schedulers’ types and also with respect to all the other MAC schedulers mentioned in the references.

PT

Keywords-component: LTE-A, MAC, Scheduler, QoS, Pareto Optimal, Min-Max, NSGAII

CE

I. INTRODUCTION 3rd Generation Partnership Project (3GPP) body has come up with high-speed broadband wireless network specification (Release-10 and beyond), which is known as Long Term Evolution Advanced (LTE-A) architecture [1]. According to [1], the LTE-network deployment architecture is shown in Fig1. It consist of multiple base-stations (referred as eNodeB) giving coverage to some designated geographical area. These eNodeBs are then connected to another LTE-A network element termed as Evolved Packet Core (EPC), via an IP based packet-switched network, sometime termed as IP backhaul network. Typical throughput of this IP backhaul network per eNodeB is 1Gbps. All the application servers like video streaming; audio streaming, etc. are connected to EPC via internet. The user equipment (UE) once establishes the session with eNodeB and EPC, will then be able to connect to the application servers.

AC

Figure 1: LTE-A Deployment and eNodeB Data Plane Architecture

In Fig-1, we also depict a two practical architecture showcasing how a PHY layer gets connected to the MAC protocol layer. In one of the architecture, there is a tight coupling between MAC protocol layer with the PHY layer (i.e. eNodeB-1) and in other architecture, there is a loose coupling between the layers using an internal IP interface (i.e. eNodeB2).The above eNodeB-2 design of architecture is becoming popular due to cloud based radio access networks (RANs) for LTE-A. In this paper we will analyze the impact of eNodeB-2 design architecture.

1

ACCEPTED MANUSCRIPT

The 3GPP LTE-A specifications [1] mandates researchers to develop innovative MAC scheduler module inside MAC protocol layer to select the right user(s) and its queue(s) every transmission time interval (TTI) in order to meet the Quality of Service (QoS) requirements. To the best of our knowledge, all the MAC schedulers studied so far (Refer-Section II) follows single objective function with single decision constraints. For LTE-A type of system, these MAC scheduler implementations will result in high packet drop, low radio resource utilization and degradation of user throughput for a high capacity scenario.

A.

LTE MAC SCHEDULER ALGORITHM AS DESCRIBED IN PRIOR ARTS:

The MAC broadband telecom scheduling has evolved steadily from Universal Mobile Telecommunications System (UMTS) till LTE-A architecture. Some of the early LTE-A MAC scheduling methods are as follows; (a)User selection based on the application type and its priority[3],[4],[5],(b)Scheduling decision based on user’s application buffer size status[6].All the above MAC scheduling decision is based on single constraints which are simple decision and fails to provide required LTE-A QoS. These scheduling methods require lot of enhancement to incorporate decisions logic like the impacts of hybrid automatic repeat request (HARQ), radio condition and queuing delays.

CR IP T

With respect to above limitations, in this paper, for first time, proposed a novel scheduling technique, which will simultaneously address multi objective and with multiple operation constraints (as the issues mentioned in Section II) and maintain the LTE-A QoS requirements. Our new MAC scheduler is termed as Multi-Objective QoS-aware Downlink Scheduler (MOQDS)which shall improve QoS, along with good fairness in terms of radio resources and user allocations in two levels eNodeB2 architecture. Using multi-objective optimization approach, MOQDS comes up with a scheduling scheme that maximizes the overall eNodeB’s performance. Once the user and its queue selection are complete, MOQDS’s does the radio resources’ allocation and sends the allocation information to the PHY layer for air transmission.

AN US

In recent past, QoS has been the top concern area of all the service providers due to multiple challenges in wireless network. So we analyzed quite a few of prior arts focusing on QoS based scheduling approach. Some well referred QoS based scheduling approaches are,(a)Scheduling of latency sensitive applications using the tail probability method[7],(b)Guaranteed scheduling of VoIP traffic at fixed transmission slots [8],(c)Improving QoS by efficient resource allocation techniques [9],(d)Scheduling decision based on rate and userspecific QoS requirement [10],(e)Scheduling decision based on mean opinion score for VoIP application[11]. For all the above scheduling methods, authors though improved QoS mainly for real time applications, but its fails to address the effect of different channel conditions and mixed traffic condition.

ED

M

While analyzing the result, MOQDS illustrates that there is a significant performance improvements compared with other well known scheduling techniques like, modified largest weighted delay first (MLWDF), exponential rule proportional fairness (EXP-PF) and LogRule schedulers under different traffic models and ITU channel model like PedA, VehA[2]. According to the results, MOQDS is able to achieve minimum 50% reduction in packet drop rate and minimum three times increased cell throughput compared with the other MAC schedulers mentioned in the references.

SCHEDULING IN BROADBAND CELLULAR SYSTEMS: PRIORWORK,MAJOR CHALLENGES AND PROPOSED CONTRIBUTIONS

AC

II.

CE

PT

The rest of the paper has been organized as follows. Section II, presents the overview of relevant scheduling algorithms in the prior work and the major challenges. Section III presents the MOQDS algorithm and its analysis. Section IV, defines the scheduler performance metrics for evaluating the MOQDS. Section V, defines the system simulation parameters. Section VI, reveals the results and corresponding discussion. Section VII, concludes the present work with futuristic possible studies.

This section is divided into two sub-sections. In the first sub-section i.e., sub-section-A, we have analyzed several different LTE-A MAC scheduler algorithm as described in prior works and summarized the yet to be solved major drawbacks. In sub-section-B, we have described proposed MOQDS contribution in resolving those drawbacks as well as increasing the performance. While analyzing these prior works, to reduce confusion, we assume that the application, traffic, service flows and flows are the same entities and having the same definition.

As demand surged towards giving guaranteed QoS, we saw that some prior arts started moving away from single constraint to multiple constraints for their scheduling objective functions. The scheduling logic used for multi constraints QoS schedulers are as follows; (a) Scheduling based on user throughput and delay as described in [12],[13],(b)QoS aware scheduling using highest value of the head of line packet delay and timeaveraged user throughput as described in [14],[15],(c)Scheduling logic based on mean delay-optimal scheduling policies and throughput values of both real-time flows (LogRule)[16],(d) Scheduling algorithm based on Virtual Token Modified Largest Weighted Delay First (VT-MLWDF), Modified Largest Weighted Delay First(MLWDF) and Exponential PF (EXP-PF)for different classes of traffic such as Video, VoIP and best-effort[17],(e) Scheduling logic which is based out of logical channel priority and user’s channel quality[18].After analyzing these prior arts, it was concluded that the above scheduler algorithms were designed for guaranteeing QoS without considering multiple applications per user basis. Also, the schedulers mentioned above area single monolithic block, and handling more than one constraint is not a practical approach due to increased processing time. In recent past, many authors started the multi levels LTE-A MAC scheduling approach to overcome the above drawback in previous paragraph. Some of multi levels scheduling approaches are as follows: (a)In [19], the scheduling frameworks consists of an upper level scheduler which computes data size to be allocated to the user using frame level scheduler. The lower level scheduler does the radio resource allocation based proportionate fairness rule. (b)In[20], a two

2

ACCEPTED MANUSCRIPT

level scheduling logic consists of QoS handling for each application type at upper level along with priority handling mechanism at lower level.(c)In [21] and[22], the two level scheduling algorithms consist of an upper level scheduler designed using discrete-time linear control theory for application queue handling. The lower level scheduler does the radio resource allocation based on proportionate fairness rule. Though the above prior arts deals with Two-Level scheduling concept, but proposes to use single constraint per level basis. This is again a simplistic approach in designing scheduler for meeting LTE-A QoS requirements.

6.

CR IP T

Thus to develop a true QoS aware scheduling algorithm for LTE-A network, we still need to answer some fundamental questions based on practical Fig-1 eNodeB-2 architecture which are not handled in the entire prior art schedulers. 1. According to Fig-1, eNodeB-2 has two main external interfaces (marked as A and D in Fig-1). Most of the MAC scheduler described in the prior art deals with a monolithic scheduling block, handling simultaneously for these two interfaces. The scheduling decision was not very effective as it deals with less number of constraints which eventually leads to QoS degradation. Attempts are been made in defining multi-level scheduling decision, but that also does not take into consideration of the right constraints at the each of these levels. So, the challenge is how to split the monolithic scheduling block into multilevels catering to all the interfaces of the eNodeB along with right constraints. 2. According to Fig-1, users of the LTE-A system will have multiple logical channels associated with different applications. So the problem is how these multiple logical channels per users affect the QoS of eNodeB. What are the mechanisms by which we can select the appropriate logical channels among user, which will improve the overall QoS. 3. According to Fig-1, the S1UP is an IP interface. Hence, packet arrives to the eNodeB system at random rate due to complex network deployment topology. This leads to a jitter at the S1UP interface. Due to this introduction of the jitter at the S1UP interface, user’s QoS gets affected as the scheduler cannot schedule packet regularly at the air interface. So the problem is how to design a QoS scheduler which is aware of the all S1UP interface jitter. 4. According to Fig-1, once the packets enter the eNodeB system, there are multiple protocol layers (e.g. GTPU, PDCP,RLC) that processes the packets before it lands to the users’ queues at MAC layer. This introduces a processing latency which varies significantly as the call load increases in the eNodeB system. Due to this processing latency, each user’s MAC queues do not get replenished deterministically thus affecting the user QoS due to non-availability of packets in the queue. So the problem is how to design a QoS aware scheduler which considers the user’s MAC queue replenished rate. 5. According to Fig-1, we have shown two data plane architecture demonstrating how the PHY layer gets connected to the MAC protocol layer. With respect to the

second architecture (i.e. eNodeB-2), under peak load scenario there is a communication delay experienced between the MAC and PHY interface. Due to this communication delay, the user scheduled packets will reach late to the PHY layer and misses the TTI slot causing drop in user throughput. So the problem is how an efficient MAC scheduler should be designed which considers the communication delay between MAC-PHY interfaces so that the scheduled packets does not miss the TTI. According to 3GPP specifications [1], if the user receives corrupt packet(s) it sends negative acknowledgment to the eNodeB and the scheduler executes retransmission procedure known as Hybrid Automatic Repeat Request [1]or HARQ. The entire prior art, does not consider or allocate any extra priority to do this HARQ retransmission. So the problem that needs to be solved is how to do priority re-balancing among user during HARQ retransmission scenario.

AN US

With respect to above open challenges, we have proposed following novel ideas to solve. B.

PROPOSED CONTRIBUTIONS:

In this paper, our detailed novel contributions with respect to a true LTE-A QoS aware scheduler are described below:

ED

M

1.

CE

PT

2.

AC

3.

4.

3

To the best of our knowledge, this work is unique in developing a true QoS aware scheduling algorithm for LTE-A network where the MAC scheduling is done every TTI that too in two levels. The idea of two levels scheduling was conceptualized after analyzing the eNodeB’s two external interfaces as shown in Fig-1 (marked as A and D). In our proposed MOQDS, each level operates independently on each of the above mentioned eNodeB’s interfaces with a well defined objective function and multiple constraints. To achieve the LTE-A QoS goal, MOQDS defines an upper MAC scheduler operating at S1UP interface and the lower MAC scheduler operating at the air interface. The main goal of the upper MAC scheduler objective function is to do an optimal users’ queue selection based on its multi constraints. Similarly, lower MAC scheduler objective function is to do an optimal users’ selection based on its own multi constraints. To address the stringent LTE-A QoS goal, the novelty of the paper is to identify some new operational constraints which were deemed necessary. These new operational constraints are as follows: (a) S1UP interface jitter, (b) MAC queue replenishment rate per logical channel per user, (c) Throughput deficit of each logical channel per user, (d) MAC to PHY layer communication delay variation. Along with the above mentioned operational constraints, we also considered and integrated some the well-known legacy constraints such as: (a)MAC queue priority of each logical channel associated with each user. (b) MAC queuing delay for each logical channel(s) associated with

ACCEPTED MANUSCRIPT

6.

Symbol and definition

:S1UP incoming jitter

: The normalized S1UP incoming jitter experienced by the user’s queue at S1UP interface : The normalized queue replenishment rate for the user’s queue arising due to processing latency at data plane protocol layers : The normalized MAC queue priority for the user’s queue as defined by [1] : The normalized MAC queuing delay for the user’s queue experienced due to non scheduling to user’s queue.

: MAC queue replenishment rate : MAC queue priority : MAC queuing delay

Thus using the above novel methods, MOQDS is able to achieve (a) highest QoS (i.e. highest cell and user throughput, low packet drop rate and minimum latency with respect to above well referred MACs), (b) fairness among radio resource allocation, (c) ability to identify right logical channels per users per TTI (d) handle simultaneous users’ transmission per TTI (e) user priority adjustment during retransmission scenario.

Where, i varies from

Eq. No. (1) (2)

(3) (4)

and k varies from

AN US

Upper MAC Scheduler Objective Function: According to MOQDS, the upper MAC scheduler objective function for the user’s queue at the scheduling time i.e. is considered as a time varying entity and is dependent on S1UPincoming jitter experienced by queue ( ), MAC queue replenishment rate ( ), MAC queue priority of the user queues ( , MAC queuing delay ( )as represented in Eqn. (5a).

PROPOSED MOQDS ALGORITHM FORMULATION AND SOLUTION In this section, we will describe the complete algorithm and analytical model of the objective functions associated with each level along with their multiple operational constraints. This section is divided into three parts. The sub-section A deals with objective function formulation and its solution approach using multi-objective method. The sub-section B deals with objective function formulation and its solution approach using heuristic method. Lastly, sub-section C shall deal with the physical block allocation strategy. For the algorithm formulation, we assume that there are N numbers of users and each user has M numbers of queues.

ED

M

III.

Subject to: : : : :

OBJECTIVE FUNCTION FORMULATION AND SOLUTION USING MULTI-OBJECTIVE METHODS

In this sub-section, we will present the different operational constraints associated with the two objective functions operating at each eNodeB’s interfaces. After that, we will use different solution methodologies to solve these two objective functions operating to come up with the scheduling recommendation.

(5a)

Thus, the first optimization problem can be formulated as given in Eqn.(5b). By maximizing the upper MAC scheduler objective function, we try to find out the optimal set of users queue for which we can attain the desired QoS. :

PT

III.A

Constraint Name

CR IP T

5.

each user (c) Achievable spectral efficiency. This constraint is nothing but the spectral efficiency associated with user’s channel quality indicator (CQI) [1] at an instant of time. (d) Previous transmission spectral efficiency. This constraint is nothing but the final spectral efficiency achieved by the scheduler at previous scheduling instant of time. Every TTI, these two objective functions of MOQDS does independent scheduling and comes up with an optimal user and its MAC queue recommendation for maximizing the overall eNodeB’s QoS and fairness. Once the user and its queue selection are complete, MOQDS’s does priority re-balancing taking the HARQ retransmission into consideration.

∑

∑

(5b) (5c)

CE

Where, is defined as the maximum S1UP incoming jitter experienced in the IP backhaul by the users queues, is defined as the maximum MAC queue replenishment rate (QRR)for the users queues, is defined as the maximum value of users MAC priority supported by the eNodeB, is defined as the maximum MAC queuing delay for the users queues. As seen in Eqn.(5a), it is a generic definition which we will modify with a proper mathematical equation. To achieve this task, we first individually define a sub-objective function which is dependent on upper MAC operational constraints. We would likely to define four sub-objective functions, which will then be combined to form a single upper MAC scheduler objective function. After we analyzed the physical behavior of the S1UP incoming jitter and MAC queuing delay, we concluded that

AC

III.A.1 UPPER MAC SCHEDULER- OPERATIONAL CONSTRAINTS AND OBJECTIVE FUNCTION FORMULATION

To attain the goal of selecting an optimal set of users queue in every TTI, in this section we would first define the upper MAC scheduler’s operational constraints and then derive an objective function which will depict the scheduling behavior using these operational constraints. Thus, the Upper MAC scheduler operational constraints for user’s queue at instant of time is given in Table 1: Upper MAC scheduler operational constraints: Table 1: Upper MAC scheduler operational constraints

4

ACCEPTED MANUSCRIPT

their behavior can be emulated as an exponential based utility function. Thus, we defined two independent sub-objective functions for the user’s queue at instant of timewhich is shown in Eqn.(5d) and Eqn.(5e). The basic definition of the exponential based utility function is taken from [26].

time is given in Table 2: Lower MAC scheduler operational constraints

Table 2: Lower MAC scheduler operational constraints

(5e) In similar way, we define the MAC queue replenishment rate as a sigmoid function [27] and form 3rd sub-objective functions for the user’s queue at the scheduling time which is shown in Eqn.(5f). ⁄

Eq. No.

: The normalized achievable spectral efficiency for the user at the scheduling time

(6)

:

:

Throughput deficit

(5f)

Where, is the utility convergence factor such that . Lastly, we define the fourth sub-objective function for the user’s queue at the scheduling time which is linearly dependent on the MAC queue priority as shown in Eqn.(5g).

Proposition 1:At scheduling time, the upper MAC scheduler objective function for the user’s queue after combining the four sub-objective functions is given in Eqn.(5h). (5h) =

M

)

ED

Proof: Observe that, the upper MAC level scheduler objective function should not be biased and dependent on any particular operational constraints. Thus, analyzing the system functionalities as given in Fig-1, the upper MAC scheduler objective function for user’s queue should be dependent on the below two points. 1. Upper MAC scheduler shall try to schedule the user queue which is of higher priority and experiencing high MAC queuing delay. Hence, upper MAC scheduler objective function should be directly proportional to the sub-objective function related to user queue priority and MAC queuing delay. 2. Upper MAC scheduler shall restrict the scheduling of the MAC queues which are experiencing high S1UP incoming jitter and high MAC queue replenishment rate. Hence, upper MAC scheduler objective function is inversely proportional to the sub-objective function related to S1UP incoming jitter and MAC queue replenishment rate.

The

normalized throughput deficit for the user at the scheduling time. Where is defined as the maximum bit rate configured, is defined as the average user throughput : The normalized MAC to PHY outgoing delay variation for user at the scheduling time : The normalized previous allocated spectral efficiency at scheduling time, which is dependent on the average user throughput i.e.

AN US

: MAC to physical layer delay variation :Previous allocated spectral efficiency

(5g)

(

Symbol and definition

CR IP T

(5d)

Constraint Name :Achievable spectral efficiency

(7)

(8) (9)

Lower-MAC Scheduler Objective Function: As per the MOQDS, the lower-MAC scheduler objective function for the user’s at the scheduling time i.e. is considered as a time varying entity and it is dependent on achievable spectral efficiency , throughput deficit , outgoing jitter and previous allocated spectral efficiency as presented in Eqn. (10a). =

(

,

, )

(10a)

CE

PT

Thus, the second optimization problem can be formulated as given in Eqn.(10b).By maximizing the lower-MAC scheduler objective function, we try to find out optimal users’ selection set for which we can attain the desired QoS. :

AC

Subject to: : : : :

∑

(10b) (10c)

Where, is maximum spectral efficiency supported by the eNodeB, is defined as the cell throughput degradation, is defined as the maximum in the MAC to PHY transmission delay variation. As seen in Eqn.(10a), it is a generic definition which we will now modify with a proper mathematical equation. To achieve this task, we first individually define a sub-objective function which is dependent on lower MAC operational constraints.

III.A.2 LOWER-MACSCHEDULER - OBJECTIVE FUNCTION AND ITS CONSTRAINTS

To attain the goal of selecting the optimal set of user in every TTI, in this section we would first define the lower level MAC scheduler’s operational constraints and then derive an objective function which will depict the scheduling behavior using these operational constraints. Thus, the Lower MAC Scheduler operational constraints for user at instant of 5

ACCEPTED MANUSCRIPT

We would likely to define four sub-objective functions, which will then be combined to form a single upper MAC scheduler objective function. The first sub-objective function for the lower MAC scheduler is formed using simple logarithmic based objectivefunctions with throughput deficit as its variable parameter as shown in Eqn.(10d).

constraint values. Hence, the solution approach is nothing but a multi-objective optimization problem. So, we reorient the MOQDS as a multi-objective optimization problem [23]which is achieved by combining the two objective functions and then a solution of these two objective functions are being done either by Min-Max principle [24]and by Non-Dominated Sorting Genetic Algorithm(NSGAII)[25] principle. Thus, the LTE MAC scheduler which uses the MinMax solution principle is named as MOQDS-MM and MOQDS-NSGAII which uses the NSGAII solution principle. In general, all the above solution swill lead to a user and queue pair selection in order to attain the maximum QoS. From now onwards we drop the time variable notation (t) for all the constraints variables to simplify the presentation.

(10d)

CR IP T

The second sub-objective function is formed using linear functions with MAC to PHY outgoing delay variation as its variable parameter which is shown in Eqn.(10e).

III.A.3.1. SOLUTION USING MIN-MAX PRINCIPLE

(10e)

In this approach, we first linearly combine the two objective function as given in Eqn. (5b) and Eqn. (10b) using a linear weight as given in Eqn. (11a).

The third sub-objective function is formed using linear variation of the ratio of allocated spectral efficiency to the achievable spectral efficiency as its variable parameter which is shown in Eqn.(10f). ⁄

AN US

(10f)

Proposition2. At scheduling time, the lower MAC scheduler objective function for the user’s queue after combining the three sub-objective functions is given in Eqn.(10g). (10g) =

ED

M

Proof: Observe that, the lower MAC level scheduler objective function should not be biased and dependent on any particular operational constraints. Thus, analyzing the system functionalities as given in Fig-1, the lower MAC scheduler objective function for user’s should be dependent on the below two points. 1. Lower MAC level scheduler shall try to schedule the user more often who are experiencing higher throughput deficit and experiencing high MAC to PHY outgoing delay variation. Hence, lower MAC scheduler objective function should be directly proportional to the subobjective function related to throughput deficit and MAC to PHY outgoing delay variation. 2. Lower MAC level scheduler shall restrict the scheduling of the user whose allocated spectral efficiency matches the achievable spectral efficiency. Hence, lower MAC scheduler objective function is inversely proportional to the sub-objective function related to ratio of allocated spectral efficiency to the achievable spectral efficiency.

=

FOR

+

=

PT

APPROACH

0= =

(11b) From the above we can conclude that each objective function assumes the same objective value, irrespective of the weights and .We now convert the Eqn. (11b) as individual maximization problem as given in Eqn. (11c) and simplifying the solution by equating each of the each objective function to zero. (11c) and

CE

AC

III.A.3 SOLUTION FUNCTION

(11a) = Where is combination of two field state vector which are as defined as [ , , , and as defined as [ , , .As seen in Eqn.(11a), the choice of the linear weight can bias the either towards or towards . This will lead to infinite number of solution space. In order to converge to a single solution space, we consider Min-Max principle [24]. The Min-Max principle requires that we find to maximize ; then we can denote the value of the corresponding state variable by . Using the approach given in [24] we find out which is maximized at , which is given below:

MULTI-OBJECTIVE

Subject to: as given in Eqn. (5c) and as given in Eqn. (10c). To solve Eqn.(11c), we could have used the traditional methods such as brute force or branch-and-bound to search the optimal solutions space. All these methods cannot be used because it is very time consuming in giving the solution. Since the LTE-A scheduling decision has to happen in real time and

According to earlier sections, we see that MOQDS is being devised as two objective functions, namely the upper and lower MAC scheduler. As described in the earlier sections, each of the two objective functions has different objective and its goals. More often, these two objective functions may not be able to reach a common and optimal solution under the same 6

ACCEPTED MANUSCRIPT

that to in 1 milli-second; we use the Karush-Kuhn-Tucker (KKT) conditions to solve the above optimization problem. Hence, the KKT’s stationarity conditions are given in Eqn. (11d) and Eqn.(11e). (11d) ( ) ( )

(

)

(

) )

) ) )=0

=0

All the constraints as defined in Eqn. (4b-4e) and Eqn. (10b10e) supports the KKT’s primal feasibility and all the multiplier supports the dual feasibility as given in Eqn. (11h) (11h)

ED

) ) ) (

) )

} [

]

Since, MOQDS-MM scheduler is executing at every TTI of one milli-second, it is impossible to find every time an optimal user and its relevant queue for QoS maximization. Thus, the solution path has to be a sub-optimal approach. In this approach, we try to find the best case user and its queue every TTI by choosing maximum Euclidian distance of the objective function using current value of the constraint variables and optimal value as shown in Eqn. (14).

1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14.

AC

=0

=0

=0

Using the gradient descent method, the KKT multipliers can be updated as given in Eqn.(12b). (12b) [ ( )] (

{

}

{

}

Algorithm-1: MOQDS-MM algorithm

=0

[

)]

(14)

The complete algorithmic approach of MOQDS-MM is described in Algorithm-1.

=0

CE

(

(

{

PT

(

(12a)

=0

)

)

(

M

) (

(

[ [

)]

[

By taking the derivative of Eqn.(11d), and Eqn. (11e), we get the is given in Eqn. (12a).

(

(

AN US

(11g)

(

[ [ [

)]

{ }and ( Where [ are positive step sizes. We use the iterative optimization technique to solve Eqn.(12a) and (12b). First, we take the updated KKT multipliers from Eqn.(12b) and substitute to Eqn.(12a). Then from the suboptimal solution of Eqn.(12a), we again update back to Eqn.(12b). This step is repeated till the iteration converges and this point we declare Eqn.(12a) and Eqn.(12b) has reached the optimal solution. This optimal solution we term as utopia solutions where a user and its chosen queue in the eNodeB system has attained an optimal value of S1UP incoming jitter (i.e. , MAC queue replenishment rate (i.e. , MAC priority (i.e. , MAC queuing delay (i.e. , achievable spectral efficiency (i.e. , throughput deficit (i.e. , MAC to physical layer delay variation (i.e. and previous allocated spectral efficiency (i.e. . Let, the optimal value of the each objective function as given in Eqn.(13). { } , (13)

(11e)

The KKT’s complementary slackness is given in Eqn. (11f) and (11g) (11f) ( ) ( ( (

(

CR IP T

(

[

)]

7

[

]

[ {

}

ACCEPTED MANUSCRIPT

USING GENETIC

9. [ 10. { } 11. (b) [ 12. (b1)Determine the Non-Dominated Sort:

NONALGORITHM

After analyzing the Eqn. (5b) and (10b) along with the constraints defined in Eqn. (5c) and Eqn. (10c), we conclude that it is a problem of multi-objective optimization, which will lead to multiple decision boundaries. Since the objective functions are non-trivial, there could exit multiple sub-optimal solution that will simultaneously optimize each of Eqn. (5b) and (10b). Thus, we will try to find out the non-dominated Pareto optimal solutions as per multi-objective function given in Eqn. (15).

{ } { }

{

22.

(15) Subject to: where is combination of two constraint vector which are = [ , , , & =[ , ,

(16)

AN US (

M

{

}

34.(b3)Selection [25]: 35. 36. 37. III.B

(17)

CE

Algorithm-2: MOQDS-NSGAII algorithm

based

on

crowded

comparison

OBJECTIVE FUNCTION FORMULATION AND SOLUTION USING HEURISTIC METHOD

=

AC

3.

)

In this section, we will now consider a totally different approach of formulating the two objective functions and solution approach using a simple heuristic method. Thus, MOQDS which uses the Heuristic method is denoted as MOQDS-H. The above solutions will also lead to a user and queue pair selection in order to attain the maximum cell and user throughout along with user fairness. So, in this method we will redefine Eqn. (5a) and Eqn. (10a) that is the upper MAC and lower MAC objective functions respectively to Eqn. (18a) and Eqn. (18b).

The complete Pareto solution approach of MOQDS-NSGAII is described in Algorithm-2.

-Initialize

{ }

29.

ED

PT

}

,

{ } 27. (b2)Calculate the Crowding Distance [25]:

As per the theory, a point is Pareto optimal if there does not exist another point, such that , and .Thus, in this section we solve the multi-objective functions as defined in Eqn. (15) using the genetic algorithms (GA) approach. As the GA works with population of points, it was our natural choice to use it so as to find out the multi-objective optimization solution. There are varieties of GA methods that can be used for determining Pareto optimal solution; we have taken into account the Nondominated Sorting Genetic Algorithm II [25] principle. The reference [25]is widely cited paper for calculating Pareto optimal with fast convergence. Since, MOQDS-NSGAII scheduler is executing at every TTI of 1 milli-second, it would generate a Pareto optimal solution contour. Now to choose the best case user and its queue at every TTI Pareto optimal solution contour, we apply the maximum Euclidian distance of the contour’s current value at instant of time as shown in Eqn. (17). {

}

CR IP T

III.A.3.2. SOLUTION DOMINATEDSORTING PRINCIPLE

=

(18a) (18b)

As per Eqn. (18a), the upper MAC objective functions is an adhoc numeric addition of the four distinct constraints of the upper MAC namely, , , , as given in Eqns. (1), (2), (3), (4) respectively. Similarly according to Eqn.(18b), the lower MAC objective functions is also an adhoc numeric addition of the four distinct constraints namely , , as given in Eqns.(6), (7), (8), and (9).Since, MOQDS-H scheduler is executing at every TTI of one milli-second, we first calculate objective function value of Eqn.(18a) for each

{ } 4. 5. { } 6. Loop: 7. (a) Start with Population Initialization [25]: 8. [ ]

8

ACCEPTED MANUSCRIPT

user’s all the queues at the scheduling time.After that, we take only select the highest value of Eqn. (18a) from all its ‖ as given in M queues and then add it in the sorted list ‖ the Eqn. (19a).

[ 4. 5. 6.

‖

(19a) ‖ =arg. { } Once the upper MAC objective function accomplishes in calculating the capability of each of the user queues and then sorting the user queues, this becomes the starting point of the lower MAC scheduler. For each of the user at the scheduling time, we only select the highest value of the objective function Eqn. (18b) from all the users. And then ‖ as given in the Eqn. (19b) added in the sorted list ‖ ‖ ‖ = arg (19b) The complete algorithmic approach of MOQDS-His described in Algorithm-3.

IV.

CR IP T

AN US PHYSICAL RESOURCE STRATEGY

BLOCK

M

III.C

SCHEDULER PERFORMANCE METRICS

To evaluate the performance of different version of MOQDS with respect to all other types of scheduler as described in Reference, we define four scheduler performance metrics. These four metrics are the user throughput, the overall cell throughput, packet drop rate and Jain-FairnessIndex. Brief descriptions of these metrics are given in following paragraphs. As scheduler is ready for transmission, the user’s acknowledged (ACKed) throughput at time instant is given in Eqn. (21). This is the first scheduler performance metric. The first parameter is the defined as total acknowledged bits until instant of time. The second parameter is the time difference between current time and start time in milli-seconds. (21) = /

Algorithm-3: MOQDS-H algorithm approach 1. 2. 3. 4. Loop: 5. 6. 7. 8. 9. 10. 11. 12.

]

ALLOCATION

CE

PT

ED

According to [1], the scheduled users are allocated specific number of OFDMA sub-carriers for a predetermined amount of time in order to do air transmission. These OFDMA subcarriers are referred to as physical resource blocks (PRBs) in the specification [1]. PRBs thus have both a time and frequency dimension. So the resource allocation strategy means how to distribute the physical resource block (PRB) [1]in a fair and opportunistic way among the scheduled users. The PRB allocation continues till the PRB list is not exhausted. Taking the retransmission [1]concept into consideration, the lower MAC scheduler objective function as defined in Eqn. (10a) and (18b) i.e. is re-considered taking the Boolean value of HARQ retransmission status. If “retx” is the packet transmission feedback from user, then Eqn. (10a) get modifies as given in Eqn. (20). (20) =[ Thus, every TTI if the user at scheduling time reports a negative acknowledgement, then the user’s lower MAC objective function turns out to be unity, and bypass all the logic of user selection as per section-III-A and B. The final PRB allocation algorithm is given in below steps.

At the end of every scheduling cycle, scheduler finally calculates the overall cell throughput. This is defined as the addition of all the scheduled user(s) ACKed throughput present in the eNodeB. So the overall cell throughput at scheduling time is presented in Eqn. (22) which is the second scheduler performance metric. (22) =∑ Where, N is defined as the total number of users in a cell and n is the total number of user out of N scheduled in scheduling time. The packet drop rate (PDR) is third scheduler performance metric which is dependent on two attributes. The first attributes is the which is measured as total packet dropped by the scheduler either due to packet delay crossing the maximum delay threshold for that application or the packet retransmission counter has crossed the maximum allowed threshold value. The second attribute is which is defined as the total acknowledgement packets from the UE. Thus, the PDRis presented the Eqn. (23).

AC

PDR=

(23)

Lastly, we measure the fairness performance efficiency of the MOQDS using Jain Fairness Index [29](JFI) which is the fourth scheduler performance metric. Assuming, there are N users and is the throughput for the user, and then the Jain Fairness value is defined in Eqn. (24). 〈∑ 〉 (24) )= ∑

Algorithm-4: PRB allocation algorithm

9

ACCEPTED MANUSCRIPT

as described in Section IV for all the wellknown schedulersas mentioned in Introduction. We have presented the results broadly into three main categories. SectionVI-A has presented the cell and users’acked throughput results analysis of different types of schedulers. SectionVI-B has presented the packet drop rate results analysis and in section VI-C hasanalyzes the Jain fairness index. During our discussion, stating MOQDS would mean that it consists of all the variants namely Heuristic, Min-Max and NSGAII. Also, we synonymously use users and UEs.

The result ranges from (worst case) to 1 (best case), and it is maximum when all users receive the same allocation. V.

SYSTEM SIMULATION AND PARAMETERS

In this paper, we have considered an open source based LTE-A system simulation tool which has been configured as per Table-1.All the simulation parameters are been taken mainly 3GPP specification [1] as well as from references [30]. The architecture and design changes of the tool were carefully done so that it emulates Fig-1. The simulation tool emulates multiple user equipments, multiple eNodeBs, lightweight evolved packet core and lastly the two different application servers. One of the application servers emulates VoIP traffic and other one emulates Video streaming. The data throughput of each of these application servers are given in Table-I. From the evolved packet core to the eNodeB, i.e. the S1UP interface, we have introduced the interface jitter. In similar way, we introduced the MAC-PHY interface jitter. In the simulation tool, we then added the complete MOQDS (Heuristic, Min-Max and NSGAII) scheduling logic as well as other scheduling logic from the reference like EXP-PF, MLWDF, and Log Rule. At the end of the simulation, we generate graphs for all the four different scheduling performance metrics. For ease of simulation time, we have taken a cell layout of 17 omni-directional rather than tri-sector per eNodeB configuration.

Max Delay (ms)

Max Number of Re-Transmission

AC

CQI

Number of UE Number of Queue per User Traffic Generator

eNodeB and UE antenna gains Antenna Configuration

VI.

CR IP T

M

ED

CE

UE Mobility

Value 2GHz 5MHz, 20MHz 25, 100 10% 1ms 17 Omni directional 1Km L=128.1 + 37.6log10(R) BTS Ant Ht = 32 m, MS = 1.5m 0 mean, 8dB PedA, Veh30 43dBm -174 dBm/Hz 3 km/hr , 30km/hr, User moves till Cell Edge and Return 100 (VoIPCall) or 400 (Video Streaming) 3 Full bandwidth and periodic reporting scheme. Measured period: 2 ms 10 and 50 per serving base station 2 (VoIP and Video Streaming Type) Large Size Packet generator – 5Mbps Small Size Packet generator – 64kbps 0 dBi 1x1

PT

Path Loss Model Log Normal Shadowing Channel Model eNodeB Transmit Power White Noise Power

CELL AND USER ACKED THROUGHPUT ANALYSIS

In this section, we describe in detail the analysis of the cell and users’ ACKed throughput occurrences with respect to 50 users’ in a cell, channel bandwidth of 5MHz as well as 20MHzand channel model as VehA. For the complete simulation duration, each user is having two queues, namely VoIP and Video streaming. We will present the results for cell and users’ ACKed throughput taking the Video streaming only into consideration and compare the result with different schedulerslike EXP_PF, MLWDF, and Log Rule. In Fig.2, we specifically study the entire range of schedulers’ users’ ACKed throughput of Video steaming with 50 cell users’, eNodeB system bandwidth of 5MHz and the channel model as VehA. The X-axis of the Fig.2 is Signal-toNoise Ratio (SNR) variations in dB and the Y-axis is the users’ ACKed throughput in Kbps. The simulation tool derives the SNR from the wideband CQI [1]. It is observed that MOQDS users’ ACKed throughput is more compared with the other scheduler types like EXP-PF, MLWDF, and Log Rule. This is due to the fact that, MOQDS scheduler’s lower MAC objective function constraints is able to consider the convert the achievable spectral efficiency from CQI as per Eqn.(6), calculate the user throughput deficit using Eqn.(7), takes into consideration of previous spectral efficiency using Eqn.(9) in every TTI, and it is able to choose the right user more often compared with the other scheduler types, resulting in increased the user’s throughput .

AN US

Table 3: Considered parameters for Scheduling Parameters Name Carrier frequency System bandwidth Physical Resource Blocks Target Block Error RateBLER Transmission Time Interval(TTI) Cellular Layout Cell Radius

A.

Figure 2: User ACKed Throughput - 50UEs ; 5MHz ; VehA - Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDSMinMax, MOQDS-NSGAII

RESULTS AND DISCUSSIONS

In this section, we present the simulation results of MOQDS (Heuristic, Min-Max and NSGAII) with respect to analytical model as described in Section III, and present the comparison with respect to the scheduler performance metrics

Observing Fig.2, it can be seen that the users’ ACKed throughput for Video streaming increases as the SNR value 10

ACCEPTED MANUSCRIPT

CR IP T

increases. This is due to the fact that MOQDS uses Eqn.(6), (7) and (9) to start the PRB allocation of higher transport block size compared with other scheduler types like EXP-PF, MLWDF, and Log Rule. In general, as the transport block size increases, the instantaneous user’s throughput as presented in Eqn. (21) also increases, which results in increase of users’ ACKed throughput. For the lower SNR values, the PRB allocation controls the optimum size of the transport block so as to minimize the multiple retransmissions. Still, there is drop of users’ ACKed throughput at lower SNR values. This is because; the packet transmission goes into retransmission mode. By multiple retransmissions ensures that successful delivery of the user packets, but the time to deliver the packet increases. This reduces the user’s ACKed throughput. In higher values of SNR, the effect of packet retransmission is not that significant. Hence, we can conclude that the MOQDS schedulers have users’ ACKed throughput variation more compared with the other scheduler types for all the ranges of SNR. In Fig.3, we kept all the system parameters same as Fig.2, except eNodeB system bandwidth is increased to 20MHz from 5MHz. From Fig.3, we can conclude that the MOQDS schedulers have users’ ACKed throughput is more compared with the other scheduler types for all the ranges of SNR. The only difference between Figs.2 and 3, is that the magnitude of the users’ ACKed throughput in Fig.3 is more than Fig.2, which is due to fact that the number of PRBs in 20MHz is 4 times than that of 5MHz resulting higher size of transport blocks. In Figs.4-5, we study entire schedulers’ number of user scheduled per TTI of 50 cell users’ scenario, under VehA channel conditions with eNodeB’s system bandwidth as 5MHz and 20MHz respectively. The X-axis of the Figs.4-5 captures the user scheduled per TTI and Y-axis captures the cumulative probability distribution of the user scheduled per TTI occurrences.

Figure 4: User scheduled per TTI - 50UEs – 5MHz - VehA - Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDSMinMax, MOQDS-NSGAII

AC

CE

PT

ED

M

AN US

From Fig.4, it can be concluded that MOQDS is able to scheduled more than 10 users per TTI, which is among highest in all the other scheduler types like EXP-PF, MLWDF, and Log Rule. From Fig.5, it can also be concluded that MOQDS is able to scheduled more than 25 users per TTI, which is among highest in all the other scheduler types. As described in section III-C, MOQDS scheduler uses efficient PRB allocation rule, which in turn increases the number of users scheduled per transmission time interval. As we increase the eNodeB’s system bandwidth from 5MHz (as in Fig.4) to 20MHz (as in Fig.5), the number of PRBs also increases resulting more users per TTI getting scheduled by MOQDS compared with the other scheduler types. The cell throughout variation for all the scheduler types MOQDS, EXP-PF, MLWDF and LogRule are shown in Figs.6-7. The X-axis of all the Figs.6-7 captures the cell throughput and Y-axis captures the cumulative probability distribution of the cell throughput occurrences.

Figure 5: User scheduled per TTI - 50UEs – 20MHz - VehA – Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDSMinMax, MOQDS-NSGAII

Figure 3: User Acked Throughput - 50UEs – 20MHz - VehA - Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDSMinMax, MOQDS-NSGAII

11

Figure 6: Cell Throughput - 50UEs - 5MHz - VehA – Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDS-MinMax MOQDS-NSGAII

CR IP T

ACCEPTED MANUSCRIPT

Figure 7: Cell Throughput - 50UEs - 20MHz - VehA–Video Streaming, EXP-PF, MLWDF, LogRule, MOQDS-Heuristic, MOQDS-MinMax, MOQDS-NSGAII

In Fig.6, we specifically study the cell throughput occurrences of the Video streaming for 50 cell users, eNodeB system bandwidth set as 5MHz and the channel model as VehA. It is being observed that the cell throughput for MOQDS schedulers is more than the other scheduler types like EXP-PF, MLWDF, and LogRule. To understand this variation, we need to refer to cell throughput as considered in Eqn. (22). It is being shown that the cell throughput is dependent on the users’ ACKed throughput according to Eqn. (21) and total number of scheduled users per TTI out of total cell users i.e. N. As seen from Figs.2-3, users’ ACKed throughput of MOQDS scheduler is more than the other scheduler types. Also, from Figs.6-7 it can be concluded that MOQDS schedules more users per TTI out of total cell users, i.e. N. Thus, the combined effect of the more users’ ACKed throughput variations along with the high number of users scheduled per TTI makes MOQDS scheduler’s cell throughput of the Video streaming higher than the other scheduler types. In Fig.7 we studied the effects the cell throughput variation of Video streaming due to the change in eNodeB’s system bandwidth as 20MHz and keeping all the other parameters like number of users, channel model same as Fig.6. As seen in Fig.7, due to the increase of eNodeB system bandwidth to 20MHz, the number of PRBs that the scheduler uses for resource allocation also increases as shown in Table-I. This is because more users are getting scheduled in every TTI. MOQDS scheduler uses efficient PRB allocation technique which causes in increasing the cell throughput compared with the other scheduler types. From Fig.6, comparing the 95 percentile value of cell throughput value of all other scheduler types (referred as point-A in the Fig.6) with MOQDS (referred as point-P,Q,R in the Fig.6),it can be concluded that MOQDS is minimum 2.7 times more improved than all other scheduler types. From Fig.7, it can be concluded that MOQDS is minimum 7 times more improved of all scheduler types. We also found that apart from the scheduler types like EXP-PF, MLWDF, and LogRule, MOQDS performance with respect to user or cell throughput is better than the entire MAC schedulers like FLS, FLS-M-EXP and VT-M-LWDF as mentioned in the references.

B.

PACKET DROP RATE (PDR) ANALYSIS

AC

CE

PT

ED

M

AN US

The PDR variation of different scheduler algorithms under different system parameters are presented in Figs. 8-9. For each bar-graph figure, the X-axis represents the system variables like the total number of users, the channel bandwidth and user traffic class. The Y-axis represents the PDR value against the above sets of system variables. We have also analyzed the PDR performance of different scheduler algorithms by changing the channel model from PedA to VehA. For the set of bar-graphs [a1, c1] and[b1,d1] shown in Fig8, the PDR values of MOQDS are appreciably low compared with other scheduler like EXP-PF, MLWDF and LogRule for VoIP as well as Video streaming. This is because of the efficient handling of the user queue using upper MAC objective function with its right set of constraints as described in Eqns.(1),(2),(3) and (4). The other schedulers like EXP-PF, MLWDF and LogRule does not have the provision of handling these conditions, which results into high PDR values. As seen from the set of bar-graph [a1] and [c1] shown in Fig8, as the channel bandwidth increases from 5MHz to 20MHz, the number of PRB also increases proportionately. The scheduler types like EXP-PF, MLWDF, and Log Rule able to reduce the queuing delay, but not that significantly. MOQDS take the advantage of Eqn. (1), (2), (3), (4) to alter the upper MAC objective functions accordingly with which it is able to handle the queuing delay, along with HARQ retransmissions effectively. Thus, the PDR of MOQDS is again appreciably low compared with other scheduler like EXP-PF, MLWDF and LogRule. As we compare the [a1,a2], [b1,b2], [c1,c2] and [d1,d2] of Fig-8, due to the increase in the number of users, packet in each user’s queue wait for longer duration of time before getting the PRB from the scheduler. If the queuing delay crosses maximum allowed delay time, all the other scheduler types like EXP-PF, MLWDF, and LogRule discards it causing high PDR value. But MOQDS adjust the upper MAC objective functions using Eqn. (4); so that the packet gets scheduled before the delay crosses maximum allowed delay time.

12

ACCEPTED MANUSCRIPT

CR IP T

like the total number of users, the channel bandwidth and user traffic class. The Y-axis represents the JFI values for the above sets of system parameters. For the Fig-10 and 11, we analyze the JFI performance for the PedA and VehA channel model respectively.

Figure 8: Cell PDR - PedA–Video Streaming and VoIP

The effect of Veh-A channel model on PDR is being studied in Fig.9. As we compare all the set of bar-graphs between Fig-8 and 9, the channel model causes short-term fading; delay spread which results in increasing the high packet retransmission. None of the scheduler types like EXPPF, MLWDF, and Log Rule handles the retransmission effectively causing high PDR value. But MOQDS takes into the retransmission effect and adjust the lower MAC objective function during PRB allocation phase according to Eqn. (20). This reduces the PDR compared with other scheduler types like EXP-PF, MLWDF, and Log Rule. Comparing with Fig. 89, PDR of MOQDS is well within 4%, which shows a better channel variation performance compared with other scheduler types like EXP-PF, MLWDF, and Log Rule. Thus, it can be concluded that the MOQDS supports minimum 50% improvement of PDR handling for the same system bandwidth compared with EXP-PF, MLWDF, and LogRule and also with other MAC schedulers like FLS, FLS-M-EXP and VT-MLWDF as mentioned in the references.

AC

CE

PT

ED

M

AN US

Figure 10: Jain Fairness Index - PedA - Streaming and VoIP

Figure 9: Cell PDR - VehA–Video Streaming and VoIP

C.

JAIN FAIRNESS INDEX (JFI) ANALYSIS

The JFI variation of different scheduler algorithms under different system parameters are presented in Figs. 10-11. For each bar-graph figure, the X-axis represents system variables 13

For the set of bar-graphs [a1, c1] and [b1, d1] shown in Fig-10, the JFI values of MOQDS are higher compared with other scheduler like EXP-PF, MLWDF and LogRule for all the category of traffic class. This is because MOQDS take the advantage of scheduling multiple users per TTI as described in Section-III-C, so as to alter the objective function correctly for QoS along with maintaining the proportional allocation of PRB among the users. This helps in maintaining a high fairness index compared with the other scheduler types as EXP-PF, MLWDF, and Log Rule. Similar behavior is observed for the set of bar-graphs, i.e. [a1, c1] and [b1, d1] as shown in Fig-11. As we compare the [a1,a2], [b1,b2], [c1,c2] and [d1,d2] of Fig-11, due to the increase in the cell users, packet in user’s queue wait for longer duration of time before getting the PRB from the scheduler. Thus maintaining the allocation fairness of PRB among the users’ becomes a challenge and MOQDS overcome this challenge by optimizing upper MAC objective function through Eqn. (3) and (4). After selecting the right user(s) and its right queue, MOQDS then apply the PRB allocation technique as described in Section-III-C. Similar behavior is observed for the set of bar-graphs [a1, c1] and [b1, d1] as shown in Fig-11.

ACCEPTED MANUSCRIPT

has been achieved with respect to other schedulers types like MLWDF, EXP-PF and LogRule and also with respect to all the other MAC schedulers mentioned in the references. This is significant improvement, which will help the telecommunication/ISP service providers to maximize their revenues and QoS. In the future work, we will consider the 2x2 and 4x4 MIMO models to boost cell throughput for Video streaming users. REFERENCES

[2] [3]

Figure 11: Jain Fairness Index - VehA - Streaming and VoIP

As we compare the set of bar-graphs of Fig-10 and 11, MOQDS is able to maintain the fairness during retransmission by doing the priority recalculation using Eqn. (20). But other scheduler types like EXP_PF, MLWDF, and Log Rule fail to maintain fairness as compared to MOQDS. Thus, with respect to Video streaming scenario MOQDS is better by 10% compared with EXP_PF, MLWDF, Log Rule and other MAC schedulers like FLS, FLS-M-EXP, VT-M-LWDF as mentioned in the references. Similarly, with respect to Video streaming MOQDS is better by 50% compared with EXP-PF, MLWDF, Log Rule and other MAC schedulers like FLS, FLS-M-EXP, VT-M-LWDF as mentioned in the references. From Fig.10-11, it can be concluded that for 5MHz system bandwidth, MOQDS able to handle 10 cell users with fairness measure close to unity for Video streaming and VoIP.

AN US

[4]

3GPP TS 23.401, 24.301, 29.212 , 36.300, 36.211, 36.212, 36.213, 36.321, 36.302, 36.304, 36.306, Specification (FDD)" - (Release 8). Small Cell Forum Release 5.1 – SCF159 ITU-R M.1225 Guidelines for Evaluation of Radio Transmission Technologies for IMT-R Akyildiz, H.A, Istanbul, Turkey ; Akkuzu, B. ; H kelek, I. ; Cirpan, H.A., “LTE downlink scheduler with reconfigurable traffic prioritization”, Communications and Networking (BlackSeaCom), 2014 IEEE International Black Sea Conference on, Year and date: 27-30 May 2014, Page(s):58 – 62 Sundari, K.G. ; Dept. of Electron. &Commun. Eng., Pondicherry Eng. Coll., Pondicherry, India ; Jayanthi, K. ; Gunasundari, R. “Prioritized dynamic multi traffic scheduler for downlink in LTE”, Electronics and Communication Systems (ICECS), 2015 2nd International Conference on, Year and date: 26-27 Feb. 2015,Page(s):1546 – 1550 Sunggu Choi, Kyungkoo Jun, Yeonseung Shin, Seokhoon Kang, ByoungjoChois - “MAC Scheduling Scheme for voip Traffic Service in 3G LTE”, IEEE Xplore, Vehicular Technology Conference, 2007. VTC2007 fall. 2007 IEEE 66th Publication Year: 2007 , Page(s): 1441- 1445 Guoqing Li and HuiLiu,“Dynamic Resource Allocation with Finite Buffer Constraint in Broadband OFDMA Networks”,Wireless Communications and Networking, 2003. WCNC 2003. 2003 IEEE, Year: 2003, Volume: 2, Pages: 1037 - 1042 vol.2 N. Chen and S. Jordan, “Downlink scheduling with probabilistic guarantees on short-term average throughputs,” in Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC ’08), pp. 1865–1870, Las Vegas, Nev, USA, March-April 2008 D. Jiang, H.Wang, E. Malkamaki, and E. Tuomaala, “Principle and performance of semi-persistent scheduling for voip in LTE system,” in Proceedings of the International Conference on Wireless Communications, Networking and Mobile Computing (WiCOM ’07), pp. 2861–2864, September 2007 Badri Nayak, Dr. A.M. Prasad, Mohamed Niaz M, “Application-Specific and QoS-Aware Scheduling for Wireless Systems”, International Journal of Innovative Research in Computer and Communication Engineering, Vol. 3, Issue 8, August 2015 Chao He and Richard D. Gitlin, “Application-Specific and QoS-Aware Scheduling for Wireless Systems,” IEEE 25th International Symposium on Personal, Indoor and Mobile Radio Communications, September 2-5, 2014 Chao He ; Dept. of Electr. Eng., Univ. of South Florida, Tampa, FL, USA ; Gitlin, R.D. “User-specific QoS aware scheduling and implementation in wireless systems”, Year and date: 15-17 April 2015, Page(s):1 – 7 Soni, K.,Tyagi, A. “A Suboptimal QoS Aware Multiuser Scheduling for 3GPP LTE Network”, Second International Conference on Advances in Computing and Communication Engineering, Year/date;1-2 May 2015, Page(s): 40 - 44 Geetanjal, D. Jayaramaiah, "A Downlink Scheduling Technique in LTE to Enhance QoS for Multimedia Services", International Journal of Engineering Research & Technology (IJERT), Vol. 3 Issue 4, April – 2014 B. Bojovic, N. Baldo, "A new Channel and QoS Aware Scheduler to enhance the capacity of Voice over LTE systems", Proceedings of 11th International Multi-Conference on Systems, Signals & Devices (SSD’14), Castelldefels, 11-14 February 2014, Castelldefels (Spain) Najem N. Sirhan, Gregory L. Heileman, Christopher C. Lamb, Ricardo Piro-Rael, “QoS-Based Performance Evaluation of Channel-

CR IP T

[1]

[6]

[7]

M

VII.

[5]

CONCLUSIONS AND FURTHER WORK

ED

In this paper, we started our discussion in describing different prior art schedulers with their lacuna and the identified new system challenges of LTE-A scheduler. To address these challenges, to the best of our knowledge, for first time, we proposed a novel concept of the Two-level scheduling where each of level operates independently on each of the eNodeB’s interfaces with a well defined objective function and multiple constraints. During the analysis, we have identified all the possible constraints that could exist in an eNodeB system. These objective functions are then solved using Min-Max principle and NSGAII principles. Meanwhile we also showed how to formulate a more adhoc basis multi objective function and solution using heuristic methods. In all the different versions of the MOQDS, our main aim was to user(s) and queue(s) pair selection. We also demonstrated that, how to formulate a simple PRB allocation strategy and priority recalculation during retransmission scenario. Analyzing the various simulation results, it can be concluded that the MOQDS operating at 5MHz or 20MHz system bandwidth, with 50 cell user scenarios, and with a channel profiles of PedA and VehA, there is a significant improvement of packet drop rate and cell throughput. It was observed that there is a minimum 50% reduction in packet drop rate and minimum increase of 3 times in cell throughput

[8]

PT

[9]

CE

[10]

[11]

AC

[12]

[13]

[14]

[15]

14

ACCEPTED MANUSCRIPT

[23]

[24]

[25]

[26]

[27]

[28] [29]

[30]

CR IP T

[22]

AN US

[21]

M

[20]

ED

[19]

PT

[18]

CE

[17]

AC

[16]

Aware/QoS-Aware Scheduling Algorithm for Video application over LTE/LTE-A”, pp. 49–65,YR-2015, DOI : 10.5121/csit.2015.50705 Bilal Sadiq ; Seung Jun Baek ; Gustavo de Veciana., “Delay-Optimal Opportunistic Scheduling and Approximations: The Log Rule”, Year and date30 August 2010, Page(s):405 - 418 Nasralla, Moustafa M. ; Martini, Maria G. “A downlink scheduling approach for balancing QoS in LTE wireless networks”, Personal Indoor and Mobile Radio Communications (PIMRC), 2013 IEEE 24th International Symposium on, Year: 2013 , Page(s): 1571- 1575 Wei Hong Fu, Qing Liang Kong, Wei XinTian, Cheng Wang, Li Li Ma, "A QoS-Aware Scheduling Algorithm Based on Service Type for LTE Downlink", Applied Mechanics and Materials (Volumes 347 - 350), Instruments, Measurement, Electronics and Information Engineering, August 2013, Pages - 2468-2473 TONY TSANG, "Performance Analysis for QoS-Aware Two Layer Scheduling in LTE Network" International Journal of Emerging Trends & Technology in Computer Science (IJETTCS), Volume 2, Issue 2, March – April 2013 Nasim Ferdosian, Mohamed Othman, "Two-Level QoS-Oriented Downlink Packet Schedulers in LTE Networks: A Review", Proceedings of the First International Conference on Advanced Data and Information Engineering (DaEng-2013) Piro, G. ; Grieco, L.A. ; Boggia, G. ; Fortuna, R. ; Camarda, P. “TwoLevel Downlink Scheduling for Real-Time Multimedia Services in LTE Networks”, IEEE Transactions on Multimedia, Volume:13 , Issue: 5, Publication Year: 2011 , Page(s): 1052-1065 Piro, G. ; Grieco, L.A. ; Boggia, G. ; Camarda, P. “A two-level scheduling algorithm for QoS support in the downlink of LTE cellular networks”, Wireless Conference (EW), 2010 European, Publication Year: 2010 , Page(s): 246-253 R. T. Marler and J. S. Arora, “Survey of Multi-objective Optimization Methods for Engineering,” Structural and Multidisciplinary Optimization, vol. 26, pp. 369-395, Apr. 2004 Gennert, Michael A.; Yuille, A.L, “Determining The Optimal Weights In Multiple Objective Function Optimization”, Computer Vision., Second International Conference on, Year: 1988, Pages: 87 – 89 Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan, “A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II”, IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 6, NO. 2, APRIL 2002, Page:182-197 N. Enderle ;Bourges Telecom Red ; Velez ; France ; X. Lagrange, “User satisfaction models and scheduling algorithms for packet-switched services in UMTS”, Vehicular Technology Conference, 2003. VTC 2003-Spring. The 57th IEEE Semiannual, Year: 2003 , Page(s): 17041709 vol.3 Li Chen; Bin Wang; Li Chen; Xing Zhang; Dashing Yang, “Utilitybased resource allocation for mixed traffic in wireless networks”, Computer Communications Workshops (INFOCOM WKSHPS), 2011 IEEE Conference on, Year: 2011, Pages: 91 - 96 R,M, Corless, et al, “On the Lambert W Function”, Advances in Computational Mathematics, Vol5, 1996, Pg 329-359. Rajendra K. Jain, Dah-Ming W. Chiu, William R. Hawe, “A Quantitative measure of Fairness and Discrimination for resource allocation in shared computer system”, September 26, 1984, submitted for publication to ACM transaction on computer systems Ulil S. Zulpratita, "The Adequate LTE Downlink Scheduling Scheme for Video Streaming Services", Information Systems International Conference (ISICO), 2 – 4 December 2013

15

ACCEPTED MANUSCRIPT

Hewlett Packard, Microsoft, Motorola Research, Nokia, Govt. of India in the areas of IMS and Broadband Wireless MAC/QoS/Energy-saving, TVWS. His main areas of research interest are Wireless Access Network's MAC, QoS, Power saving and IP Multimedia Subsystems. He has more than 95 peer reviewed papers in different transactions/journals and International conferences. His 7 patents are under review. He and his wireless network team had contributed three ideas to IEEE 802.16m Broadband Wireless Standard. Dr. Das received his Ph.D. degree from the Indian Institute of Technology Kharagpur.

AUTHORs BIODATA

CR IP T

Saptarshi Chaudhuri, is currently a pursuing Ph.D from International Institute of Information Technology, Bangalore and also working as Principal Architect in Wireless R&D, TATA Power Strategic Engineering Division, Bangalore, INDIA. He holds B.S. (Physics Major), B.Tech and M.Tech from Institute of Radio Physics and Electronics, University of Calcutta and IEEE Senior Member. His research interests are Next Generation Mobile Networks, Algorithm development, Network Function Virtualization and SDN. He has several papers in IEEE conferences, filled 120 Patent Applications in area of Schedulers, RRM, SON, Data & WiFi Offload, LiFi across US, Europe, India, Japan, China, with 15 US Granted Patents and 2 European Granted Patents.

AN US

Dr. Das is Chairman of IEEE Bangalore Section; Board Member of IIIT-Bhubaneswar; Technical and Empower Committee member of e-Governance, Govt. of Karnataka; Board member IT Dept. Govt. of Odisha. He is Senior Member IEEE and Fellow of Institution of Electronics & Telecommunication Engineers (IETE). Dr. Das is recipient of Outstanding Volunteer Award-2008 IEEE Bangalore Section and Global IEEE MGA Achievement Award 2012.

AC

CE

PT

ED

M

Irfan Baig, is a M.S by Research graduate in Wireless Telecommunications from IIITB, Bangalore, INDIA, also working as Lead Researcher in Wireless R&D, TATA Power Strategic Engineering Division, Bangalore, INDIA. Current research areas include LTE Layer-2 protocol (UL & DL Schedulers), RRM and LTE-Advanced SON. Currently filed 28 patents across INDIA, EU and the US in the area of Scheduler, Admission control, ICIC and ANR.

Dr. Debabrata Das is serving as Professor and Hewlett Packard Chair at International Institute of Information Technology-Bangalore (IIITB). Before joining IIITB, he had served at G S Sanyal School of Telecommunication at IIT Kharagpur and later at Kirana Networks in New Jersey. He is Principal Investigator (PI) of a project from Ministry of Electronics and Information Technology (MeitY), Government of India on Green Broadband Wireless Network and Interworking of IoT Devices from Bell Labs Nokia Research, India. He was PI of sponsored projects from Intel, 16