 Research
 Open Access
Bufferaware adaptive resource allocation scheme in LTE transmission systems
 Ruiyi Zhu†^{1} and
 Jian Yang†^{1}Email author
https://doi.org/10.1186/s136380150398y
© Zhu and Yang. 2015
 Received: 9 October 2014
 Accepted: 26 May 2015
 Published: 20 June 2015
Abstract
Dynamic resource allocation scheme is a key component of 3GPP longterm evolution (LTE) for satisfying qualityofservice (QoS) requirement as well as improving the system throughput. In this paper, a bufferaware adaptive resource allocation scheme for LTE downlink transmission is proposed for improving the overall system throughput while guaranteeing the statistic QoS and keeping certain fairness among users. Specifically, the priorities of the users’ data queues in the base station are ranked by their remaining life time or their queue overflow probability which is estimated by applying large deviation principle. An online measurement based algorithm which requires no statistical knowledge of the network conditions uses the queue priorities to dynamically allocate the resource blocks (RBs) for avoiding buffer overflow and providing statistic QoS guarantee. The simulation results show that the proposed algorithm improves the throughput and fairness while considerably reducing the average bit loss rate.
Keywords
 Dynamic resource allocation
 LTE
 QoS
 Large deviation principle
1 Introduction
Mobile communication technologies have been developed rapidly, and switched from the third generation (3G) of mobile communication systems to the longterm evolution (LTE) systems, which aims to provide highdatarate, lowlatency, packetoptimized radioaccess, and flexible bandwidth deployments [1]. LTE system allows high flexibility in the resource allocation, which enables dynamic resource blocks (RBs) allocation among the potential users [2] and [3].
Conventional resource allocation schemes in wireless system are generally based on user’s priority [4, 5]. They are designed according to user’s channel status and QoS guarantee to maximize overall system throughput. However, providing fairness among users is another essential design consideration, although it usually sacrifices the system throughput and/or violates QoS requirements. Some resource allocation schemes based on bufferaware can improve some of these performance metrics [6]. The resource allocation problem in wireless system has been widely addressed in some literatures, but it is still challenging in LTE system to design a bufferaware resource allocation scheme for improving the performance including increasing the system throughput as large as possible, guaranteeing QoS requirements, and achieving fairness.
In this paper, we propose a bufferaware adaptive resource allocation scheme by jointly considering the user scheduling and RBs allocation to provide QoS guarantee in LTE transmission systems. In the aspect of user scheduling, considering that the finite buffer maximum size, each user’s queue priority is ranked according to its remaining life time or its queue overflow probability which is estimated by applying large deviation principle. For RBs allocation, an online measurement based algorithm for dynamically allocating RBs is proposed for adjusting the service rates of the user queues in order to provide QoS guarantee. The goal is improving the total system throughput as large as possible while subjecting to provide QoS guarantee for different users and to guarantee certain fairness.
1.1 Related work
In this paper, we consider multiuser resource allocation for the downlink in LTE systems. The scheduler at the base station is responsible for allocating resources to the different users as a function of the users’ queue priority as well as the current channel conditions. There are many prior works on this problem. The classic scheduling algorithms include Round Robin (RR) algorithm [7], Max C/I algorithm [8], and proportional fair (PF) algorithm [9]. Although many works [10–13] apply multiuser diversity in user scheduling for maximizing system throughput, the system buffer size in these schemes is assumed to be infinite, that is to say, any arriving bit can be buffered and any bit loss due to buffer overflow will not happen. This assumption may not be reasonable since the buffer size is limited in the transceivers.
Resource allocation for finite buffer space has been discussed in the literature related to the wireless network. The authors in [14] design a new LTE buffer aware scheduler to opportunistically assign RBs for video streaming applications in order to maximize the average video quality. In [15], the buffer occupancy based approach is presented to achieve video rate adaptation, while in [16], a dynamic programming framework is applied to study the buffer v s. QoS tradeoff for wireless media streaming in a single user scenario. These papers cited above mainly focus on video traffic. But the eNB in the practical situation schedules and transmits general data traffic besides video traffic.
There are several related works for packet scheduling and resource allocation in wireless data systems. In [17], M. Andrews et al. focus on how to adapt MaxWeight algorithm to the multicarrier wireless data systems, and a simple variant was introduced into the objective for reducing resource wastage. In [18], M. Realp et al. propose a resource allocation algorithm in multiuser OFDMA by considering queue and channel state information. However, these methods focus on maximizing the overall throughput by improving spectral efficiency, which may lead to unfair resource sharing among users. In fact, fairness is necessary to guarantee minimum performance of the users experiencing bad channel conditions. The bufferaware adaptive resource allocation proposed for LTE system in this paper will consider the problem of keeping certain fairness while improving the total system throughput.
Due to the limited available resource, RBs allocation aims to efficiently use the shared resource and allocate the resource in a fair manner. Naturally, there is a tradeoff between fairness and system throughput. PF algorithm has emerged as a prominent candidate since it balances between fairness and throughput. In [19], S. Lee proposes a suboptimal method, i.e., PF metric(2), which introduces the status of queues into PF metric(1). However, it is pointed out in [19] that although PF metric(2) is more responsive to the queues than PF metric(1), it incurs a reduced system throughput because its isolated RB assignment strategy may assign the RB to a user having low channel quality. Similar work related to PF scheduling in LTE systems can also be found in [20, 21] and [22]. By considering both fairness and the constraint of finite buffer space, a channeladapted and bufferaware (CABA) packet scheduling algorithm is proposed in [23]. This method defines and applies the user priority in the resource allocation for avoiding buffer overflow. However, the empirical parameters in the priority function are hard to appropriately choose. Inappropriate parameters will lead to an inaccurate user priority, which induces excessive resource allocated to the users and reduce the utilization of the system resource.
The eNodeB may have large capacity to cache traffic such as audio and video streams, but it substantially increases the delay and reduces QoS. Hence, we consider the finite buffer size and queue overflow probability in this paper. We will jointly exploit the priorities of user queues and the RBs capacity for controlling the service rate of each user data queue in the base station, instead of solely relying on any one of them. Under the constraint of finite buffer space, the proposed bufferaware adaptive resource scheduling algorithm aims at achieving three objectives: (1) keep bit loss rate as low as possible by means of taking buffer status into account, (2) improve the total system throughput as large as possible, and (3) keep certain fairness among users by means of adjusting the overflow probability.
1.2 Contributions

User scheduling: Firstly, the user scheduling scheme is considering the finite buffer size. Secondly, the scheme is depending on the users’ queue priority which is calculated by their remaining life time or their queue overflow probability. The overflow probability estimation model is derived by applying the large deviation principle [24], which incorporates both the queue fullness and its variation.

RBs allocation: An online measurementbased algorithm is further presented to adjust the service rate of the user queues, which requires no statistical knowledge of the network conditions. According to the user queues’ priorities, we control the service rate of each user queue by dynamically allocating the RBs, in order to avoid queue overflow and provide statistic QoS guarantee.

We present experimental results to show that the proposed algorithm is able to improve the total system throughput while guaranteeing certain fairness among users and providing QoS guarantee.
The rest of the paper is organized as follows: Section ‘System model and problem statement’ describes a system and channel model for resource allocation. In section ‘User priority determination scheme’, we present user priority determination algorithm, including calculating the remaining life time or queue overflow probability. Section ‘Online measurementbased algorithm for dynamic RBs allocation’ is devoted to describing the online measurementbased algorithm for dynamic resource allocation. Section ‘Performance evaluation’ provides the experimental results and performance comparisons. Finally, conclusions are drawn in Section ‘Conclusion’.
2 System model and problem statement
2.1 System model
We focus on singlecell downlink resource allocation in eNB of LTE system employing OFDMA. The implementation of adaptive resource scheduling in eNB relies on the following factors: buffer status (e.g., unoccupied buffer space and current queue length), traffic characteristics (e.g., bit arrival rate) and channel quality. We assume that eNB has perfect and instant channel information for all downlink transmissions via the feedback channel, while the channel quality is assumed stationary for the duration of each subframe, but may vary from subframe to subframe. Since the data queue of each user locates in eNB, it is natural that eNB knows the amount of each user’s data in the transmissionside buffer without additional signaling to report.
2.2 Problem statement
This paper aims for maximizing the throughput while reducing BLR and keeping certain fairness among users. In order to achieve this, we jointly consider the users’ queue priority and RBs capacity for controlling the service rate of each data queue. The users’ queue priority is based on the remaining life time or the queue overflow probability which is calculated by applying the large deviation principle. Then, according to the queue priority, we adjust the service rate of each user queue through dynamically allocating the RBs, in turn providing different transmission rate to achieve statistic QoS guarantee.
where γ _{ k,n } is the signaltointerferenceplusnoiseratio (SINR) for the kth user on the nth SB.
where ξ is a given threshold value of the fairness deviation. Equation (8) indicates that each SB can be only assigned to one user during the slot t. Equation (9) indicates that the difference between 1 and the system fairness should be kept less than ξ.
The resource allocation problem (7–9) is complicated and intractable to obtain the optimal solution by exhaustive search. Here, we propose a bufferaware adaptive resource allocation scheme which considers the different priorities of user queues and RB capacity, in order to achieve a better performance tradeoff of throughput, fairness and average BLR.
3 User priority determination scheme
In order to improve the total system throughput and QoS for different users while keeping certain fairness, we need to determine the users’ queue priority for deriving an online measurement based resource allocation. The overflow probability estimation model is derived by applying large deviation principle, which requires no statistical knowledge of the network conditions. Then, we rank the users’ queue priority by their remaining life time or their queue overflow probability.
3.1 Estimation model for the queue overflow probability
The arrival rate of the incoming bits depends on the service type, while the service rate depends on the resource allocation policy as well as the wireless channel conditions which are timevarying in nature. Hence, the arrival process and the service process are independent of each other. Our aim is to control all user queues in such a way that the service demands of the data in each queue could be satisfied. Moreover, the resulted scheme should be robust to the variations of the arrival and service processes.
Let I _{ k }(t)=A _{ k }(t)−V _{ k }(t), where I _{ k }(t)∈{−m _{ V },⋯,0,1,⋯,m _{ A }}, and let \({\pi _{i}^{k}}=P(I_{k}(t)=i)\) denote the corresponding kth user queuelength variation probability distribution. Since A _{ k }(t) is determined by the bit arrival number during the slot t and V _{ k }(t) is determined by the bit number served during the slot t, their difference I _{ k }(t) characterizes the mismatch between the bit service rate and the bit arrival rate of the kth user queue during the slot t. I _{ k }(t)<0 implies that the bit service rate is higher than the arrival rate in the tth slot, while I _{ k }(t)>0 implies that the bit service rate cannot satisfy the bit arrival. Due to the timevarying number of bit arrivals and the state of SBs, the polarity of the sequence I _{ k }(t)(t=1,2,…) may change frequently between negative and positive.
where T is called prediction interval.
The term \(\frac {\sum _{i=1}^{T} I_{k}(t+i)}{T}\) in (16) is determined by the bits departure or the resource allocation, while g _{ k } is determined by the current length of the kth user queue. Since the queue overflow probability indicates the mismatch between the resource and the traffic, we can dynamically rank the users’ queue priority based on the value of \(P_{k_{\textit {overflow}}}^{t+T}\). The larger value of \(P_{k_{\textit {overflow}}}^{t+T}\) means that queue overflow is more likely to occur and the corresponding user queue should have the higher priority of resource allocation, thus reducing the bit loss rate and satisfying QoS requirement. This is why the proposed method jointly considers RB capacity and the queue priority.
In theory, the overflow probability estimate becomes more accurate as T increases. Hence, the value of T should be sufficiently large. However, owing to the rapid exponential decay of the overflow probability estimate with T, we can set T to a moderate value for the sake of acquiring an accurate overflow probability estimate. The experimental results in the section of ‘Performance evaluation’ demonstrate that T≥60 is appropriate.
In the next section, we show how to online estimate the overflow probability based on (20).
3.2 Online estimation of the queue overflow probability
According to (20), estimating the overflow probability requires the values of g _{ k }, c _{ k }, and \({\pi _{i}^{k}}\). It is easy to calculate g _{ k } according to (14). However, we have to estimate c _{ k } and \({\pi _{i}^{k}}\) because there is no prior knowledge about I _{ k }(t). Therefore, the historical observations are utilized to estimate these parameters by applying a sliding windowbased method.
Suppose the observed sequence is given by {I _{1},I _{2},I _{3},⋯ }. The sliding window covers the T _{ s } most recent entries in this sequence, which is slid over this sequence. For the nth window, the observation vector is denoted by \(\phantom {\dot {i}\!}W_{n}=[I_{n}, I_{n1}, I_{n2}, \cdots I_{nT_{s}+1}]\).
Following the similar steps in [32], we can apply the large deviation principle to analyze the confidence interval of c _{ k }.
where the parameter ρ∈[0,1]. If ρ approaches to 1, the value of \(\hat {\pi _{i}^{k}}(t)\) largely depends on the past estimation, while if ρ=0, \(\hat {\pi _{i}^{k}}(t)\) totally depends on the current estimate \( \hat {u_{i}^{k}}(t)\). According to Gardner’s report [33], ρ∈[0.7,0.9] is usually recommended.
The above steps assist us to derive the online measurementbased method to estimate the queue overflow probability \(P_{k_{\textit {overflow}}}^{t+T}\) based on (20) in the (t+T) slot, by setting T to a moderate value in a practical application. The experimental results show that T≥60 is appropriate.
3.3 User priority determination algorithm
In the case of \(\hat c_{k}\geq g_{k}\), the average growth length of the kth user queue in each slot, \(\hat c_{k}\), is higher than the achievable average growth length of the kth user queue per slot, g _{ k }, in the forthcoming T slots. This implies that if keeping the current queue configuration unchanged with the bit service rate V _{ k }(t), after T slots, the queue will be more likely to have an overflow situation. Therefore, in this scenario, we should improve the bit service rate to prolong the remaining life time. In this paper, remaining life time R _{ k }(t) can be calculated by (5) to rank the queue priority of resource allocation.
However, in the case of \(\hat c_{k}< g_{k}\), the average growth length of the kth user queue in each slot, \(\hat c_{k}\), is lower than the achievable average growth length of the kth user queue, g _{ k }, in the forthcoming T slots. But this does not necessarily imply that no overflow will happen in the future T slots, since \(\hat c_{k}\) is the average growth per slot which cannot characterize the specific queue length growth in a single time slot. Hence, the queue overflow might still occur. Since we have \(g_{k}> \hat c_{k}\), the queue overflow remains a rare event, and the queue overflow probability in the (t+T) slot, \(P_{k_{\textit {overflow}}}^{t+T}\), can be approximated by (20).
The bufferaware priority determination algorithm determines a priority value for each user, where the user in the case of \(\hat c_{k}\geq g_{k}\) is more emergent than in the case of \(\hat c_{k}< g_{k}\). The smallest value of R _{ k }(t) indicates the highest priority of the kth user. The value in ascending order represents that the users’ priority is from high to low. The smaller value is \(P_{k_{\textit {overflow}}}^{t+T}\), the lower priority is the kth user. The value in descending order indicates that the users’ priority is from high to low. According to the user queues’ different priorities, in the next section, we show how to dynamically allocate the RBs to adjust the service rate for each user queue for improving the system throughput subject to providing QoS guarantee while keeping a certain fairness.
4 Online measurementbased algorithm for dynamic RBs allocation
In this section, we will present the proposed online estimation based dynamic service rate control algorithm, which relies on a strategy of mitigating the overflow probability or extending the remaining life time.
 1.In the proposed bufferaware resource allocation scheme, at the slot t we seek the user$$ k_{1}=\arg \min_{k\in \Phi }\lbrace R_{k}(t) \rbrace. $$(25)
 2.The SB having the maximum SINR can be obtained by$$ n_{1}=\arg \max_{\textit{n}\in\Omega} \lbrace \gamma_{n,k_{1}}\rbrace. $$(26)
Then, we can calculate the transmission rate \(r_{k_{1}}^{n_{1}}(t)\) according to (6).
 3.
If \(A_{k_{1}}(t)>r_{k_{1}}^{n_{1}}(t)\), it means that allocating SB is not enough to transmit the bits in the buffer for the most emergent user queue. Let Ω=Ω∖{n _{1}} (which means removing the element n _{1} from the set Ω), then we choose the SB \(n_{2}=\arg \max _{\textit {n}\in \Omega } \lbrace \gamma _{n,k_{1}} \rbrace \) and calculate the transmission rate \(r_{k_{1}}^{n_{2}}(t)\). Compare the value of \(A_{k_{1}}(t)\) with the value of \(r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)\). If \(A_{k_{1}}(t) \leq r_{k_{1}}^{n_{1}} (t)+r_{k_{1}}^{n_{2}}(t)\), execute the step 4. Otherwise, let Ω=Ω∖{n _{2}}. Choose the SB \(n_{3}=\arg \max _{\textit {n}\in \Omega } \lbrace \gamma _{n,k_{1}} \rbrace \) and calculate the transmission rate \(r_{k_{1}}^{n_{3}}(t)\). Compare the value of \(A_{k_{1}}(t)\) with the value of \(r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)+ r_{k_{1}}^{n_{3}}(t)\), and repeat the above procedure until \(A_{k_{1}}(t) \leq r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)+ \cdots + r_{k_{1}}^{n_{m}}(t)\).
 4.If \(r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)+ \cdots + r_{k_{1}}^{n_{m}}(t) \leq Q_{k_{1}}(t)+A_{k_{1}}(t) \), the bit number transmitted in the slot t can be calculated as$$ \begin{aligned} V_{k_{1}}(t)=r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)+ \cdots + r_{k_{1}}^{n_{m}}(t). \end{aligned} $$(27)Otherwise \(r_{k_{1}}^{n_{1}}(t)+r_{k_{1}}^{n_{2}}(t)+ \cdots + r_{k_{1}}^{n_{m}}(t) > Q_{k_{1}}(t)+ A_{k_{1}}(t) \), the bit number transmitted in the slot t can be calculated as$$ \begin{aligned} V_{k_{1}}(t)=Q_{k_{1}}(t)+A_{k_{1}}(t). \end{aligned} $$(28)
 5.
Then, let Φ=Φ∖{k _{1}}, Ω=Ω∖{n _{ m }}, we seek the user k _{2}= arg mink∈Φ{R _{ k }(t)}, by repeating the procedures 2, 3, and 4, and allocate several SBs for transmitting the data of the k _{2}th user. Repeat the procedure 5 until all the users which have the value of R _{ k }(t) have been allocated with the SBs.
 6.After that, we further allocate SBs to the users who have the value of \(P_{k_{\textit {overflow}}}^{t+T}\). We choose the user$$ k_{l}=\arg \max_{k\in \Phi }\left\lbrace P_{k_{overflow}}^{t+T} \right\rbrace. $$(29)
and repeat the similar procedures 2, 3, 4, and 5 to allocate the resource and schedule users until Ω=∅.
 7.If Ω≠∅ and Φ=∅, it means that there are RBs which have not be used. In order to make the best utilization of RBs, we choose the users who have V _{ k }(t)<Q _{ k }(t)+A _{ k }(t) and constitute a new user set \(\bar {\Phi }\). We seek the user$$ k_{w}=\arg \max_{k\in \bar{\Phi}}\lbrace Q_{k}(t)+A_{k}(t) V_{k}(t) \rbrace. $$(30)
 8.Then the remaining SB with the maximum SINR can be obtained via$$ n_{w}=\arg \max_{\textit{n}\in\Omega} \lbrace \gamma_{n,k_{w}}\rbrace. $$(31)
According to (6), we can calculate the transmission rate \(r_{k_{w}}^{n_{w}}(t)\).
 9.
If \(Q_{k_{w}}(t)+A_{k_{w}}(t) V_{k_{w}}(t)>r_{k_{w}}^{n_{w}}(t)\), it means that the number of allocated SB is not enough to transmit the remaining bits in the buffer. Let Ω=Ω∖{n _{ w }}, we choose the SB \(n_{w_{1}}=\arg \max _{\textit {n}\in \Omega } \lbrace \gamma _{n,k_{w}} \rbrace \) and calculate the transmission rate \(r_{k_{w}}^{n_{w_{1}}}(t)\). Compare the value of \(Q_{k_{w}}(t)+A_{k_{w}}(t) V_{k_{w}}(t)\) with the value of \(r_{k_{w}}^{n_{w}}(t)+r_{k_{w}}^{n_{w_{1}}}(t)\phantom {\dot {i}\!}\), if \(Q_{k_{w}}(t)+A_{k_{w}}(t) V_{k_{w}}(t) \leq r_{k_{w}}^{n_{w}}(t)+r_{k_{w}}^{n_{w_{1}}}(t)\phantom {\dot {i}\!}\), we allocate the SBs \(\phantom {\dot {i}\!}n_{w}, n_{w_{1}}\) to the k _{ w }th user, if not, let Ω=Ω∖{n _{2}}, choose the SB \(n_{w_{3}}=\arg \max _{\textit {n}\in \Omega } \lbrace \gamma _{n,k_{w}} \rbrace \) and calculate the transmission rate \(r_{k_{w}}^{n_{w_{3}}}(t)\). Repeat the above procedure, choose the SBs by using the same method until \(Q_{k_{w}}(t)+A_{k_{w}}(t) V_{k_{w}}(t) \leq r_{k_{w}}^{n_{w}}(t)+ r_{k_{w}}^{n_{w_{1}}}(t)+ \cdots + r_{k_{w}}^{n_{w_{w}}}(t)\). Accordingly, we allocate the SBs \(\phantom {\dot {i}\!}{n_{w}, n_{w_{1}}, \cdots, n_{w_{w}}}\) to the k _{ w }th user.
 10.
Let \(\bar {\Phi }=\bar {\Phi } \backslash \lbrace k_{w} \rbrace \), Ω=Ω∖{n _{ w }}, and seek the user \( k_{w_{1}}=\arg \max _{k\in \bar {\Phi }}\lbrace Q_{k}(t)+A_{k}(t) V_{k}(t) \rbrace \). Repeat the procedures 8, 9, and 10 until Ω=∅ or \(\bar {\Phi }=\emptyset \).
 11.
After allocating SBs to the users in the slot t, we update the values of \(\hat c_{k}\) and g _{ k } corresponding to all the users in the slot t+1. Apply the user priority determination algorithm in the section III to rank the users’ queue priority again, and then repeat the above all procedure.
The algorithm operates at every beginning of the scheduling interval. The detail of the strategy is presented in Algorithm 1.
5 Performance evaluation
In this section, we characterize the performance of our online measurementbased adaptive resource allocation algorithm, and provide performance comparisons with other five algorithms, namely CABA algorithm [23], PF metric(1) algorithm, PF metric(2) algorithm [19], MaxWeightAlg(3) [17], and IHRR algorithm [18]. We first describe the simulation setup, and then the metrics used for performance evaluation are presented.
5.1 Experiment setup
We simulated a multiuser scenario, where the maximum number of communicating users was set to K=10,30,50. Here, the bit arrival rate for each user is assumed to obey the Poisson distribution with λ=50 k b i t/m s.
CQI is discretized into 15 levels which correspond to 15 different pairs of modulation choice and code rate. This implies that there are 15 possible transmission rates. A mapping between SINR ranges and CQIs is presented in [34]. The CQI values are used together with the number of allocated RBs to determine the transmission rates.
5.2 Performance metrics

Average bit loss rate: This metric indicates QoS of K users. It is defined as time average bit loss rate during a period of Δ, i.e.,$$ \bar{C_{k}}={1 \over {\Delta+1}} {\sum_{t=T_{0}}^{T_{0}+\Delta}{ D_{k}(t) \over A_{k}(t)} }. $$(32)
where D _{ k }(t) denotes the number of bit loss during the slot t for the kth user. Obviously, smaller \(\bar {C}\) is preferred.

Fairness: This metric is measured using Jain’s fairness index [29], which is widely applied for evaluating the system fairness. It is described as follows$$ F(t)=\frac{\left(\sum_{k=1}^{K} V_{k}(t)\right)^{2}}{K \sum_{k=1}^{K} {V_{k}^{2}}(t)}, $$(33)where F(t) denotes the fairness at time t. Then, the system fairness can be calculated according to$$ F={1 \over {\Delta+1}} {\sum_{t=T_{0}}^{T_{0}+\Delta} F(t)}. $$(34)

Average throughput: Our aim is improving the system throughput subject to providing QoS guarantee for different users. The larger average system throughput implies better performance.
All the simulation results were averaged over 50 independent runs.
5.3 Experimental results
5.3.1 Performance comparison for different user index
We used Matlab for implementing the simulations. The simulation model is based on the 3GPP LTE system model and it has a single cell with downlink transmission, where the number of RBs is 50, the carrier frequency is 2 GHz, and the system bandwidth is 10 MHz. Following the similar steps in [32], we can applying the large deviation principle to analyze the confidence interval of \(\hat c_{k}\).
Simulation parameters
Parameter  Setting 

Carrier frequency  2 GHz 
System bandwidth  10 MHz 
Transmission time interval  1 ms 
Subcarriers per resource block  12 
Resource block bandwidth  180 KHz 
Number of resource blocks  50 
Type of system  Single cell 
Channel model  Urban 
Simulation time  1000 TTIs 
5.3.2 Performance at different SINRs
5.3.3 Performance for different number of users
This section investigates the performance of the proposed algorithm and other benchmark algorithms for different number of users. In the simulation, the number of users K were chosen in the range [10, 50]. The other parameters and simulation settings were the same as those in Section ‘Performance comparison for different user index’.
5.3.4 Effect of prediction interval (T)
5.3.5 Effect of buffer size (Q ^{ m a x })
From Fig. 14, we can see that the average BLR decreased rapidly as the buffer size increased from (0.5×10^{4}) bit to (2×10^{4}) bit. This mean that too small buffer size is more likely to incur queue overflow and bit loss. As Q ^{ m a x } continues to increase, the average BLR reduces slowly. The reason is that larger capacity of the buffer has a lower probability of buffer overflow. Figure 14 shows that the proposed algorithm outperform the other methods in terms of average BLR. This benefits from the application of the user queue’s priority calculated by the remaining life time or queue overflow probability.
We also observe from Fig. 15 that the average system throughput for all strategies is improved by increasing the buffer size. The proposed algorithm performs better than other algorithms except for MaxWeightAlg(3). The reason is that increasing the buffer size decreases the queue overflow probabilities. The proposed algorithm chooses the appropriate RBs for the user queues according to their remaining life time or queue overflow probability, and the system throughput is improved. From Figs. 14 and 15, we can concluded that compared with other algorithms, the proposed method reduces the average BLR and improves the average system throughput as increasing the buffer size.
6 Conclusion
In this paper, we jointly consider user queue priority and the RBs capacity to develop a bufferaware adaptive resource allocation scheme in LTE transmission systems. Under the constraint of finite buffer space, the proposed scheme aims for improving both the overall system throughput and the statistic QoS while keeping certain fairness among users. We derived an analytical formula based on the large deviation principle invoked for estimating the overflow probability as a function of the buffer variance. Also, the remaining life time of a queue was defined, and its estimation model was presented. Both the queue overflow probability and remaining life time were applied to determine the queue priority. According to the queue priority, an online measurement based algorithm was proposed to schedule RBs for adjusting the service rate of the user queues. The proposed algorithm does not rely on any prior knowledge about network conditions. Numerical results show that compared to traditional scheduling schemes, the proposed algorithm has a better tradeoff among throughput, fairness, and QoS. It improves the average system throughput and keeps a better fairness among users, while reducing the average BLR. It should be pointed out that this paper considered all the traffic at the eNodeB. However, the emerging technology of SDN and middle deep packet inspection (DPI) boxes can be applied to identify the traffic. Hence, we will consider the application aware scheduling in our future work with the aid of SDN and DPI.
Declarations
Authors’ Affiliations
References
 A Toskala, H Holma, K Pajukoski, E Tiirola, in Proceedings of IEEE 17th International Symposium on Personal, Indoor and Mobile Radio Communications. UTRAN long term evolution in 3gpp (Helsinki, 2006), pp. 1–5.Google Scholar
 WY Yeo, SH Moon, JH Kim, Uplink scheduling and adjacentchannel coupling loss analysis for TDLTE deployment. Sci. World J. 2014, 1–15 (2014).View ArticleGoogle Scholar
 P Phunchongharn, E Hossain, DI Kim, Resource allocation for devicetodevice communications underlaying LTEadvanced networks. IEEE Wirel. Commun. 20(4), 91–100 (2013).View ArticleGoogle Scholar
 Y Peng, SM Armour, JP McGeehan, An investigation of dynamic subcarrier allocation in mimo–ofdma systems. IEEE Trans. Veh. Technol. 56(5), 2990–3005 (2007).View ArticleGoogle Scholar
 M Katoozian, K Navaie, H Yanikomeroglu, Utilitybased adaptive radio resource allocation in ofdm wireless networks with traffic prioritization. IEEE Trans. Wireless Commun. 8, 66–71 (2009).View ArticleGoogle Scholar
 J Huang, Z Niu, in Proceedings of IEEE Wireless Communication and Networking Conferenc (WCNC). Bufferaware and trafficdependent packet scheduling in wireless ofdm networks (Hong Kong, 2007), pp. 1554–1558.Google Scholar
 IC Wong, O Oteri, W McCoy, Optimal resource allocation in uplink SCFDMA systems. IEEE Trans. Wirel. Commun. 8(5), 2161–2165 (2009).View ArticleGoogle Scholar
 HAM Ramli, R Basukala, K Sandrasegaran, R Patachaianand, in Proceedings of IEEE Malaysia International Conference on Communications (MICC). Performance of well known packet scheduling algorithms in the downlink 3GPP LTE system (Kuala Lumpur, 2009), pp. 815–820.Google Scholar
 I Bisio, M Marchese, The concept of fairness: definitions and use in bandwidth allocation applied to satellite environment. IEEE Aerosp. Electron. Syst. Mag. 29(3), 8–14 (2014).View ArticleGoogle Scholar
 Z Zhang, Y He, EK Chong, in Proceedings of IEEE Wireless Communication and Networking Conferenc(WCNC), 2. Opportunistic downlink scheduling for multiuser OFDM systems (New Orleans, 2005), pp. 1206–1212.Google Scholar
 Z Diao, D Shen, VO Li, in Proceedings of IEEE Global Communicatios Conference(GLOBECOM), 6. CPLDPGPS scheduling algorithm in wireless OFDM systems (Dallas, 2004), pp. 3732–3736.Google Scholar
 G Song, Y Li, Utilitybased resource allocation and scheduling in OFDMbased wireless broadband networks. IEEE Commun. Mag. 43(12), 127–134 (2005).MathSciNetView ArticleGoogle Scholar
 S Ryu, B Ryu, H Seo, M Shin, in Proceedings of IEEE International Conference on Communications (ICC), 4. Urgency and efficiency based packet scheduling algorithm for OFDMA wireless system (Seoul, 2005), pp. 2779–2785.Google Scholar
 A Ahmedin, K Pandit, D Ghosal, A Ghosh, in Proceedings of the Conference on Emerging Networking EXperiments and Technologies (CoNEXT) student workshop. Content and buffer aware scheduling for video delivery over lte (Santa Barbara, 2013), pp. 43–46.Google Scholar
 TY Huang, R Johari, N McKeown, M Trunnell, M Watson, in Proceedings of the ACM Conference on SIGCOMM. A bufferbased approach to rate adaptation: evidence from a large video streaming service (Chicago, 2014), pp. 187–198.Google Scholar
 A Dua, N Bambos, in Proceedings of IEEE Global Telecommunications Conference (GLOBECOM). Buffer management for wireless media streaming (Washington, 2007), pp. 5226–5230.Google Scholar
 M Andrews, L Zhang, Scheduling algorithms for multicarrier wireless data systems. IEEE Trans. Netw. 19(2), 447–455 (2011).View ArticleGoogle Scholar
 M Realp, R Knopp, AI PerezNeira, in Proceedings of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2. Resource allocation in wideband wireless systems (Berlin, 2005), pp. 852–856.Google Scholar
 S Lee, Swapbased frequencydomain packet scheduling algorithm for smallqueue condition in OFDMA. IEEE Commun. Lett. 17, 1028–1031 (2013).View ArticleGoogle Scholar
 IF Chao, CS Chiou, in Proceedings of IEEE Wireless Communications and Networking Conference (WCNC). An enhanced proportional fair scheduling algorithm to maximize QoS traffic in downlink OFDMA systems (Shanghai, 2013), pp. 239–243.Google Scholar
 B Yang, K Niu, Z He, W Xu, Y Huang, in Proceedings of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). Improved proportional fair scheduling algorithm in LTE uplink with singleuser MIMO transmission (London, 2013), pp. 1789–1793.Google Scholar
 MR Sabagh, M Dianati, MA Imran, R Tafazolli, in Proceedings of IEEE International Conference on Communications (ICC). A heuristic energy efficient scheduling scheme for VoIP in 3GPP LTE networks (Budapest, 2013), pp. 413–418.Google Scholar
 L Yan, GY Yue, in Proceedings of IEEE International Conference on Wireless Communications, Networking and Mobile Computing (WICOM). Channeladapted and bufferaware packet scheduling in LTE wireless communication system (Dalian, 2008), pp. 1–4.Google Scholar
 JA Bucklew, Large Deviation Techniques in Decision, Simulation, and Estimation (2007).Google Scholar
 KI Pedersen, TE Kolding, F Frederiksen, IZ Kovács, D Laselva, PE Mogensen, An overview of downlink radio resource management for UTRAN longterm evolution. IEEE Commun. Mag. 47(7), 86–93 (2009).View ArticleGoogle Scholar
 3GPP, Further advancements for EUTRA physical layer aspects (release 9). 3GPP TR 36.814, (2010).Google Scholar
 F Capozzi, G Piro, LA Grieco, G Boggia, P Camarda, Downlink packet scheduling in LTE cellular networks: Key design issues and a survey. IEEE Commun. Surv. Tutorials. 15(2), 678–700 (2013).View ArticleGoogle Scholar
 Y Timner, J Pettersson, H Hannu, M Wang, I Johansson, in Proceedings of the 2014 ACM SIGCOMM Workshop on Capacity Sharing Workshop. Network assisted rate adaptation for conversational video over LTE, concept and performance evaluation (Chicago, 2014), pp. 45–50.Google Scholar
 A Bin Sediq, RH Gohary, H Yanikomeroglu, in Proceedings of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC). Optimal tradeoff between efficiency and jain’s fairness index in resource allocation (Sydney, 2012), pp. 577–583.Google Scholar
 JS Chase, DC Anderson, PN Thakar, AM Vahdat, RP Doyle, in ACM SIGOPS Operating Systems Review, 35. Managing energy and server resources in hosting centers, (2001), pp. 678–700.Google Scholar
 M Mandjes, Large Deviations for Gaussian Queues: Modelling Communication Networks (John Wiley & Sons, West Sussex, 2007).View ArticleGoogle Scholar
 C Budianu, L Tong, in Proceedings of IEEE International Conference on Acoustics, Speech and Signal (ICASSP), 2. Goodturing estimation of the number of operating sensors: a large deviations analysis (Quebec, 2004), pp. ii–1029.Google Scholar
 ES Gardner, Exponential smoothing: the state of the art. J. forecasting. 4(1), 1–28 (1985).View ArticleGoogle Scholar
 3GPP, Physical layer procedures (release 9). 3GPP TS 36.214, 6.2.0, (2010).Google Scholar
Copyright
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.