2.1 Scheduler in downlink LTE network

The packet scheduler located at the enodeB manages and schedules packets according to their service types. When many packets are waiting to be served, the scheduler should be able to arrange them according to service order, transmission rate and service time to guarantee satisfying their QoS requirement. The scheduler at the media access control (MAC) layer plays a vital role in measuring network performance. It allocates bandwidth to UE and is liable for determining how downlink and uplink channels are used by the enodeB and the UEs of a system [7].

Fig 1 illustrates the scheduler in the LTE protocol stack which aims to determine how to distribute radio resources to users [7]. The protocol stack functions comprise radio link control (RLC), radio resource control (RRC) and hybrid automatic repeat-request (HARQ). The listed functionalities are controlled by a downlink scheduler at the MAC layer. The RLC layer is located between RRC and HARQ. It can communicate with RRC using a service access point and logical channels with HARQ. This layer can reformat the packet data units into a size which can be accommodated by the MAC layer. Additionally, the RLC reorders the packet data unit when it is received out of order because of the HARQ operation executed in the MAC layer. Fig 2 shows how the scheduling process takes place at the scheduler by considering the channel status. The enodeB obtains the channel status from the feedback received from the users’ channel quality indicator (CQI). Resources are assigned to users on the basis of QoS requirement, fairness and channel status. The enodeB sends data and scheduling signalling using the scheduling results. Scheduled users receive data in light of RB allocation results, antenna selection and MCS in the downlink scheduling signalling.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Scheduler in LTE protocol stack.

https://doi.org/10.1371/journal.pone.0210310.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Downlink scheduling process.

https://doi.org/10.1371/journal.pone.0210310.g002

To achieve scheduling performance targets, the algorithm must be implemented at the enodeB located at the MAC layer [8]. The enodeB allocates a portion of network bandwidth to be shared among users using many approaches. LTE employs OFDMA as an approach to accommodating UEs with the different type of QoS requirements. Each OFDMA frame consists of 10 transmission time interval (TTIs). Each TTI can pan over 0.5 ms, which is equivalent to seven OFDM symbols in cyclic prefix and normal form configurations. The available system bandwidth is divided into sub-channels. Each sub-channel is of 180 kHz frequency and consists of 12 consecutive OFDM sub-carriers with equivalent spaces [9].

To reduce complexity and increase design flexibility, recent work focuses on a two-level framework [10]–[11]. This framework divides the radio resource technique into frequency domain (FD) and time domain (TD) schedulers working independently of each other. The TD scheduler chooses a group user request to be scheduled in the next TTI according to their QoS requirements. The FD scheduler often receives the selected group user and determines the RBs which should be allocated to each group user by using channel quality. The proposed algorithm in this work emphasises TD to provide effective utilisation of resources between users. Moreover, FD can be combined with most schedulers in the previous works to accomplish the provisioning of spectrum utilisation effectively.

2.2 Performance guarantee

In 4G cellular networks, performance guarantee means the manner by which network utilisation is efficiently expressed in term of fairness, QoS and system throughput provisioning. QoS illustrates how effectively and dependably a network can attain a certain level for real-time (RT) and non-RT (NRT) traffic. Several legacy approaches have been proposed for wireless communication networks for evaluating the performance of QoS requirements, as shown in Eqs (1)–(5), but each has its own drawback. For ease of reference, some notations used in this paper are listed in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Some notations used in this paper.

https://doi.org/10.1371/journal.pone.0210310.t001

Maximum signal interference-plus-noise-ratio (SINR) algorithm aims to maximise the throughput of the enodeB signal-to-noise ratio. In this algorithm, RBs are allocated only to UEs in good channel states, whereas UEs in bad channel states will never be given RBs [12]. However, maximum SINR algorithm utilises system resources but ignores fairness. It mathematically schedules user ith as follows: (1) where Ri(t) is the instantaneous data rate experienced by user i and k represents the number of RBs.

Round robin (RR) algorithm runs in a fixed round where active users are determined, and each user is assigned with equivalent and fixed number of time which is slotted in a cyclic manner. Users in RR are treated without any priority; hence, it is starvation-free [13]. This can be expressed mathematically as follows: (2) where t represents the current time, T_i denotes the last time a user was served and ith is the user on the kth RB.

The proportional fair (PF) algorithm was proposed to address the problems of fairness and user throughput by providing a balance between them [14]. It keeps records of average throughput for all the users among the active users in the system. For a given number of time slots, the algorithm selects and transmits users with the maximum rate and prevents starvation through the weighted factor such that users under bad channel conditions can be served within a time frame. Moreover, the PF algorithm does not consider delay-sensitive users; therefore, it is insufficient for supporting RT application with superior QoS. This algorithm can be expressed in LTE behaviour as follows: (3) where Ri(t) is defined in Eq (1) and λi(t) is the average data rate for user i.

Exponential PF rule (EXP-PF) is an algorithm developed to handle multimedia applications which can support systems in adaptive coding and modulation or time multiplexing systems. EXP-PF algorithm classifies users into best effort data services and streaming services. The best effort services use PF rule to maximise system throughput whilst fairness is achieved. Meanwhile, the streaming services using EXP rule aims to guarantee delay bound whilst fairness is maintained [4]. However, the algorithm cannot guarantee resource control “delay requirement” between RT and non-RT service users, and it maximises the throughput of the system. The algorithm can be expressed mathematically as follows: (4) where is the number of active downlink RT flows, Nrt denotes the number of active downlink RT flows, αiDHOL, i denotes delay of the first packet to be transmitted by the ith user, represents expected data-rate for the ith user at a time t on the kth RB and R⁻ⁱ denotes past average throughput achieved by the ith user.

Modified largest weighted delay first (MLWDF) is a sort of algorithm designed for the purpose of supporting multiple RT data users with diverse QoS requirements [15]. MLWDF provides priority to RT users. Users are served on the basis of the head of line (HOL) packet delay. It saves the probability of delayed packets above the discard bound under the highest permitted service data unit. This algorithm achieves low PLR with QoS requirement because it combines HOL packet delay with proportional fair properties when deciding on user priority. However, this algorithm is unfair because it offers a better QoS to users under favourable propagation situations than to those under other conditions. The following equation describes the metric used to represent MLWDF algorithm: (5) where a_i depends on α_i proportionally, D_i refers to the HOL packet delay of user i, R_i(t) and λ_i(t) is as defined in Eq (3)

Many researchers also developed their own radio resource scheduling algorithms to support RT services or provide fair transmissions among flows in LTE networks. The work in [10], proposed a scheduling framework for downlink LTE to satisfy QoS requirements defined by 3GPP specifications. The algorithm employs a channel-aware service concept to prioritise guaranteed bit rate (GBR) and non-GBR bearers from diverse QCIs using a sigmoid function. PDB is achieved by GBR bearers, whereas non-GBR bearers provide high spectral efficiency. In [16], an optimal and suboptimal scheduler was proposed to improve system performance. To maximise system performance, the scheduler assigns a scheduling block (SB) to the user for a period of 1 ms, which is equivalent to the duration of TTI. However, this scheduler does not show fairness among users by providing the same duration for SB and TTI as 1 ms.

A fair downlink scheduling algorithm for 3GPP networks was proposed in [8]. For each TTI, the algorithm uses an assignment model for the distribution of resources. The algorithm creates a balance between user throughput and fairness. A delay scheduler coupled throughput—fairness resource allocation algorithm was proposed in [17]. The algorithm categorises traffic into urgent and non-urgent ones. Using an optimisation technique, this algorithm finds the trade-off between RT, QoS and NRT throughput. It operates in two phases. In the first phase, it determines the scheduling priorities for the urgent packets which guarantee QoS; the second phase aims to provide fairness in terms of channel utilisation to the non-urgent packets. The work in [13], presented a comprehensive survey of the key design issues for resource allocation for downlink scheduling, although some of the assumptions in the work are also applicable for uplink scheduling. The theoretical exercises were covered, but the implementation was ignored in RT systems owing to difficulty and high cost. Moreover, only PF strategy was used for testing the algorithm.

A service-differentiated downlink flow scheduling (S-DFS) algorithm was proposed in [18] to avoid starvation of non-GBR flows and guarantee QoS of GBR flows considering channel quality, HOL packet delay, queue length and QCI. S-DFS employs channel quality and QCI to allow flow to bid for physical RB (PRB) to guarantee QoS to GBR flows. The algorithm uses adjustable α and β to prevent some flows from occupying a large number of PRBs. In addition, it reallocates unused PRBs to flow with high demand. However, a different load scenario is ignored in this algorithm. The work in [19] proposed a scheduling algorithm for multimedia applications. This allocates resources to RT traffic because they have a high demand for network resources using delay priority function. Traffic whose delay is near the threshold value is transmitted first. However, the algorithm achieves optimum QoS performance for RT applications whilst ignoring NRT applications.

The authors in [20] proposed a multi-level queue algorithm to optimise the throughput of the network. It divides users into multiple queues on the basis of their channel states. The users in good channel states are assigned high-priority queues, whereas those in bad channel states are allocated low-priority queues. This process improves the system throughput but ignores QoS requirements. In [21], an optimal priority-based rate-guaranteed radio resource allocation for LTE downlinks was proposed. It categorises users into priority and non-priority users. Priority users have the highest priority. Hence, resources are allocated to them firstly. The remaining resources, if any, are allocated to non-priority users. However, the algorithm uses the minimum data rate for allocating resources whilst ignoring users’ QoS requirements, thereby resulting in a decrease in system performance.

The above-mentioned approaches do not achieve all the QoS requirements for diverse classes of QoS, which have a significant role for the end user. Accordingly, an optimal and priority algorithm was proposed in [21]; using minimum data rate to allocate resources to the user. However, this scheme does not account for QoS requirements for different users which lead to performance degradation. Motivated by the above, a novel radio resource scheduling algorithm is proposed, which improve the performance degradation and guarantees QoS provisioning.

2.3 Radio resource allocation optimisation

The downlink packet scheduler must guarantee the fulfilment of the QoS requirements of cellular 4G networks and be robust to variations in channel states. Therefore, downlink scheduling in LTE networks is an NP-hard problem because the optimised scheduling process is complex and time consuming [22]. Recently, an optimal heuristic approach for multi-users in downlink LTE was proposed. The algorithm aims to maximise the total throughput of users in the network using overbooked scenarios. However, computational complexity is high [23]. A two-level scheduling algorithm which uses cooperative game theory and a virtual token mechanism was suggested in [24]. To improve the radio resource scheduling algorithm in the downlink system, the algorithm adopts a bankruptcy and the Shapley value for efficient and fair distribution of bandwidth between flow classes. However, computational complexity and weak scalability are main concerns of this approach.

Chance-constrained programming for cognitive OFDM networks was investigated in [25]. The scheme aims to maximise spectral efficiency; however, it causes interference to primary users owing to the high transmission power. In [26], a novel scheduling algorithm based on game theory and multi-criteria decision making in LTE networks was presented. The scheme improves resource allocation for smart grid applications using bankruptcy and Shapley (cooperative game theory) techniques for order performance by similarity to ideal solution. The measure drawback of this scheme is its complexity, where scheduling decision-making consumes 1 ms.

In [21], an optimal priority algorithm (hereafter called the Benchwork algorithm) was implemented for LTE downlink scheduling. In this scheme, the scheduling problem is formulated using minimum data rate and users are served on the basis of priority order. However, selecting users without considering the class range does not yield optimal value. Additionally, other LTE performance targets must be identified to achieve efficient radio resource allocation. The major contributions of this work are the introduction of a novel radio resource scheduling algorithm and provision of an optimal solution for the proposed scheme.

3.1 Problem formulation

Every QCI has a specific packet despatching behaviour that is based on the certain specification. To fulfil the QoS class-based constraints and maximise total throughput, the class bearers are allocated to a particular QCI that is based on scheduling approach while satisfying the QoS requirements. In this proposed work, it is assume that the network model has an enodeB which allocates RBs to the k number of users. Each OFDMA has 1 ms duration and contains 12 or 14 RBs of OFDM symbols. The bit data rate r_i equivalent to an allocated single RB can be expressed as follows: (6) where represents the code rate associated with the MCS, M_i denotes the size of the MCS, T_s represents the OFDM symbol duration, N_RB denotes the number of OFDM symbols in an RB (usually 12 or 14, according to the cyclic prefix or normal cyclic used) and L_e(j) represents the number of data-carrying subcarriers.

In the LTE setup, CQI is described on the basis of channel coding rate and MCS. The scheduler will determine which MCS should be imposed if the nth RB is assigned to the kth user. Moreover, the maximum CQI on all RBs of the kth user can be defined as follows: (7) where n represents the maximum CQI, CQI_{k, n} denotes the CQI of user k in the nth RBs, the user’s CQI on N RBs can be defined as CQI_{k, n} = (CQI_k,1, CQI_k,2,…, CQI_{k, N}), N represent CQI values, argmax is selection in the range [min(CQI_k),…, max(CQI_k)].

Furthermore, let CQI_{k, n}, n = 1, 2, …, N_i be a real scalar or vector sent back by the user k to show the combined channel qualities of all the subcarriers within the nth reported RB. Let q_{i, max} ∈ (1, 2, …, I) be the index of the highest-rated MCS which can be achieved by the k user for the nth RB at CQI value CQI_{k, n} i.e . Define ω_{k, n} as the resource allocation indicator for user k on the nth RB. Assume that each RB is allocated particularly to one user, i.e., if ω_{k, n} = 1 and ω_{k′, n} = 0 ∀ k′ ≠ k. We also assume that C_{k, j} represents the selection of the MCS of user k on all the RBs assigned to it in one subframe. If C_{k, j} = 1 then C_{k, j} = 0 ∀ i ≠ j. Hence, the kth user’s achievable data rate (bit/s) r_k in one subframe can be described as follows:

(8)

Therefore, the proposed algorithm, which intends to attain maximum sector throughput by considering users’ achievable data rate of each user, can be expressed as follows: (9) Subject to: (10) (11) (12) where R_k denotes the rate requirement for user k, i ∈ N = {1,…, n} represents the scheduling bearer index selected for user i. Each bearer i is allocated to a single QCI type with class {1,…,9}. PDB_k and PER_k denote the standardised PDB and PER threshold values of the equivalent QCI type, respectively, and pdb_{i, k} and per_{i, k} represents the measured PDB and PER for bearer i from QCI type k, respectively. Eq (9) denotes the objective function defining the total data rate achieved in the present TTI. Eq (10) represents the data rates needed by users to guarantee QoS requirements. Eqs (11) and (12) illustrate the computed values for PDB and PER for bearer i.

Eq (13) ensures that each RB is allocated to a single user, whereas Eq (14) sets the limitation that all RBs for a given user must use the same MCS.

(13)

(14)

The problem in Eq (9) is nonlinear because of the product ω_{k, n}, C_{k, j}. Although we can have a solution for the optimisation approaches, global optimality cannot be assured [27]. Therefore, to prevent this predicament, the problem in Eq (9) can be converted into an equivalent linear problem by introducing an auxiliary variable a_{n, k, j} = ω_{k, n}, C_{k, j}.

From Eq (13) ω_{k, n} = 1, therefore C_{k, j} × 1 = C_{k, j}. By substituting ω_{k, n}, C_{k, j} with a_{n, k, j} as given in Eq (15), the problem can be linearized as illustrated below: (15) Subject to (10), (11), and (12) (16) where R denotes a large positive real number.

Consequently, Eq (15) can be solved by the integer linear programming method [27]. Therefore, the scheduling problem for this formulated optimisation technique is the manner by which enodeB decides which user should be scheduled for each RB because users have different QoS requirements. Thus, the advantages of our proposed algorithm are maximisation of system throughput and minimisation of delay and PLR, as given in Eqs (9)–(16). Moreover, Eq (15) is an integer optimisation problem which is rendered NP-hard by the numbers of constraints and variables. Consequently, a new scheduling algorithm is proposed to minimise the complexity problem of Eq (15). Algorithm 1 further shows the pseudocode for the proposed QCI radio resource allocation algorithm for LTE downlinks.

Algorithm 1: Quality of Service Class Identifier (QCI) Radio Resource Allocation Algorithm for LTE Downlink

1 Input:

2 K ← number of user;

3 N_RB← number of RBs (25);

4 R_k← rate requirement for user k;

5 G ← feedback CQIs matrix on RB by user k;

6 P_k← priority index for user k (1 or 0);

7 W ← available number of RBs when every user was served;

8 N_k← estimated number of RBs required by each user;

9 β_CQI← channel quality indicator;

10 r_r← achievable data rate;

11 RA_p← allocate RBs to priority users;

12 RA_np← allocate RBs to non-priority users;

13 g_k← average channel gain

14 Initialization:

15 T_k = {}, k ∈ {1, 2, 3, …, K};

16 ω_{k, n} = 0, k ∈ {1, 2, 3, …, K}, n ∈ {1, 2, 3, …, N_RB};

17 r_k = 0, k ∈ {1, 2, 3, …, K};

18 Calculate g_k =

19 Calculate N_k =

20 Users are arranged in decreasing order of their data rate requirement

21 for k = 1 to K do

22 if P_k = 1 then

23  RA_p hold;

24 end

25 if N_k ≤ W then

26  allocate N_k to user k, and calculate r_k as in Eq (8)

27 end

28 select the N_k RBs of user k based on Eq (7), and insert RBs in the set T_k

29 RBs in the set T_k are aranged based on smallest CQI

30 if r_r ≥ R_k then

31  k = k + 1

32 end

33 if P_k = 0 then

34  RA_np hold;

35 end

36 if RBs still remains after allocation to non-priority users, then redistribute equally to priority users

37 end

3.2 Complexity analysis

In this section, the complexity analysis for PF, EXP-PF, MLWDF and the proposed algorithm is given by considering allocation and computational times [7]. The complexity of PF algorithm is calculated by the selection of the best metric for user n, and its scheduling complexity is given as O(log n) for each RB. MLWDF algorithm adds a weighted factor and HOL packet delay to the PF metric. However, its computational time differs from that of PF, and it chooses from several users. Hence, it has an allocation time complexity of O(log n) for each RB. Similarly, for EXP-PF algorithm, an exponential delay parameter for calculating priority metrics is computed by focusing on resource allocation cost, thereby resulting in the allocation complexity of O(log n). Furthermore, the scheduling complexity for the proposed algorithm is based on a sorting technique. Based on simple insertion sorting, it has the complexity of n², whereas for quick, merge and heap sorts, the complexity is O(log n).

Similarly, in terms of computational time, the proposed algorithm shows how time is minimised efficiently, and the computational time is calculated on the basis of packet loss by considering the initial and final times of transmission. Hence, the time complexity based on this scenario is O(1). Furthermore, assume K users with different services types. Thus, in every TTI, the time to count K users with I number of service generated is O(K*I). In this algorithm, we consider two cases for resource allocation, i.e. channel condition and QoS requirement of user traffic, which are implemented only once in the process for (2^K−1) hence resulting in the time complexity of O(I*(2^K−1). Therefore, the proposed algorithm’s computational time is computed as O(K*I)+O(I*(2^K−1).

4.1 Discussion

The results of our finding are discussed in this section. The results obtained in Section 4 show that the proposed algorithm, which achieves outstanding data throughput, outperforms all other algorithms when the network condition is stable because UE with better channel condition obtains more RBs. For video and VoIP traffic, all other algorithms significantly increase the PLR and the delay. By contrast, the proposed algorithm performs well by minimising the PLR and the delay at an acceptable fairness index with the help of the introduction of the new radio resource allocation scheme and considering user QoS requirements. Additionally, Benchwork, RR, PF, EXP-PF and MLWDF attempt to maintain fair transmission between traffic such that VoIP traffic can perform well. However, they still remarkably increase the average packet delay of large-demanded video traffic. Our proposed algorithm can result in better performance of VoIP with the help of bearer class priority and the resource allocation approach. The proposed algorithm guarantees the QoS requirement of video traffic with the delay tolerance of 130 ms, as shown in Table 2. Complexity analysis of computational time shows that the proposed algorithm has an outstanding performance, reduced computational time and merely 6.52% CPU time consumed for resource scheduling. These features allow the proposed algorithm to be a valid candidate for packet scheduling in future LTE technologies, such as 5G and IoT.

The proposed algorithm outperforms conventional scheduling algorithms and can thus be regarded as an innovation for downlink LTE networks.