QoE-driven cross-layer optimization in wireless networks addressing ...

45th Annual Simulation Symposium

QoE-driven cross-layer optimization in wireless networks addressing system efficiency and utility fairness Srisakul Thakolsri, Serdar Cokbulan, Dan Jurca, Zoran Despotovic, Wolfgang Kellerer DOCOMO Euro-Labs, Ubiquitous Networking Research Group, Munich, Germany Email: thakolsri,cokbulan,jurca,despotovic,[email protected]

concept with utility max-min allocation that corresponds to the satisfaction of each user in the system being equal regardless of the application type and the channel quality condition. They showed that the utility max-min allocation outperforms the bandwidth max-min allocation under a variety of utility functions for different application classes. In [5], Mean Opinion Score (MOS) was used as a common utility metric for userperceived quality in the cross-layer optimization. The proposed MOS-based scheme was extended by targeting at different objective function such as a modified max-min fairness that guarantees a minimum service quality to all users [6]. In [7], it has been shown that applying any of the MOS-based approaches leads to significant improvements of user perceived quality compared to other approaches including the throughput maximization. However, none of the prior art deals with the optimization of wireless resource allocation addressing multiple objectives simultaneously. In this paper, we focus on a QoE-driven cross-layer optimization for wireless video delivery that takes into account two objectives: utility maximization and utility max-min fairness. The utility is defined as a degree of user-perceived quality of the service delivered by the network operator as proposed in [5]. The first objective emphasizes achieving a maximum average perceived quality of all users which can be interpreted as how efficient the network resources are used and distributed to all users. Whereas for the second objective, its goal is to achieve a similar perceived quality among all users. It emphasizes minimizing the quality difference between the user experiencing the highest quality and another user experiencing the lowest quality. We formulate and solve the optimization problem which allocates the total system resources among the active clients so as to satisfy the chosen optimization criteria. While in perfect systems with users having good channel conditions, all applications can be served with very high quality, in constrained systems, the resource allocation must be carefully performed as a consequence of the trade-off between the system efficiency of the allocated resources and the fairness balancing among all users. When these two operation points are too far apart, the network operator may prefer to have an intermediary operation point. The major contribution of

Abstract—We address a multi-criteria Quality of Experience (QoE) driven optimization problem for multi-user wireless video delivery. The optimization is done in a cross-layer fashion by using parameterized models of application and link layer in order to determine the application data rate and the network resources that fulfill the criteria. Mean Opinion Score (MOS) is used as a unified utility metric that encompasses the userperceived quality under certain receiving conditions for the user application. We consider two partially contrary objectives, which are maximizing the total perceived quality of the whole system and maximizing the utility fairness among all users. A tuning mechanism is proposed, which allows the network operator to change dynamically its operating point of the resource allocation based on its pre-defined policy of any combinations of these two objectives. We implemented the proposed tuning algorithm in an emulated Long-Term Evolution (LTE) system and verified its feasibility through simulations.

I. I NTRODUCTION The global mobile data traffic continues to grow exponentially due to a tremendous demand for mobile video delivery. Even with mobile network upgrades, a huge ramp in mobile video traffic is still considered to be a major reason causing a network congestion problem. The mobile access networks remain a bottleneck link of mobile multimedia communication when providing mobile multimedia services to a large number of users. Hence, an efficient utilization of scarce radio network resources becomes a priority for the mobile network operation. Optimizing the wireless resource allocation to multiple clients accessing different applications at the same time, over the same wireless medium has been extensively studied over the past years. Most works are based on the Cross-Layer Optimization (CLO) scheme [1], in which key parameters of different layers are exchanged and used to perform a joint optimization with the objective of throughput maximization. Eryilmaz et al. [2] combined the throughput maximization with queue length information in scheduling to ensure fairness of resource allocation. Utility based optimization was first proposed in [3], where a concave utility function for an elastic traffic (e.g., web-browsing, file transfer) is used to capture the user satisfaction as a function of data rate. Therein, the resources are allocated so as to achieve a maximum utility of the whole system. Cao et al. [4] introduced a new fairness

978-1-4673-0040-7/11/$26.00 ©2011 IEEE

12

Fig. 1.

Target use case and network configuration considered in this paper.

this paper is to design a tuning mechanism allowing a system operator to dynamically adjust its operation point between the extreme points of maximum system efficiency and maximum fairness of perceived quality among all users. Its purpose is to provide flexibility to the network operator, and not to show whether the tuning algorithm outperforms the existing QoEbased optimization schemes as already shown in [7]. Figure 1 depicts the single cell scenario in LTE, where the limited resources are shared among users watching different videos stored at the application server on their mobile terminals. A high video quality is provided by encoding the video with a high data rate. The QoE optimizer acts as a traffic manager in the core network that makes a decision on resource allocation and rate adaptation. In our work, we assume that application utility functions are either stored in advance or sent along with the data stream. Also, the constraint of system efficiency and user fairness are previously set by the network operator prior to the optimization. Abstracted information from the base station and delivery of the target data rate from the results of optimization to the rate shaper module are done via an explicit signaling.

Fig. 2.

Relationship between data rate and signal-to-noise ratio (SNR) [8].

shown in Figure 1. Prior to the optimization, the QoE module determines an arbitrary transmission data rate at MAC layer Ri by varying the normalized fraction of resource share αi allocated to user i as follows: Ri = αi · Rmax,i , 0 ≤ αi ≤ 1, i ∈ S

(1)

where S is the set of users, S = {1, 2, · · · , N }. As a return to the estimation of Rmax,i , the LTE base station receives the optimal resource allocation αopt,i for each user i that provides the optimal data rate Ropt,i appropriate for the channel to the user i. To realize this requirement, we use the MAC-layer scheduler based on the packet-based generalized processor sharing (PGPS) strategy [9] that assigns resource proportionally to the given αi value. III. A PPLICATION LAYER MODEL

II. R ADIO LINK LAYER MODEL

In our work, we focus on the video application and derive its utility function using the MOS metric to capture the userperceived video quality, which has been originally proposed to measure voice quality perceived by the user [10]. The scaling of MOS is ranged between 1 to 4.5, which can be interpreted as from ”unacceptable” to ”excellent” quality. Due to the LTE MAC-layer retransmission mechanism, we assume that the video transmission over the wireless interface is reliable. Thus, we simplify the video utility U as a function of only transmission data rate as described below:

We adopt the long term link layer model in the LTE base station as proposed by Saul et al. [8], since it is less complex in terms of parameter estimation, parameter exchange, and performing the optimization. Hence, the network operator is flexible in placement of the QoE optimization module anywhere in its network. We estimate the maximum achievable data rate Rmax,i at the MAC layer for each user i, if all wireless resources are allocated to user i. Due to the user mobility, signal fading from the multipath effect, channel shadowing from urban obstacles, as well as the effect of noise and interferences from external sources, the wireless channel condition is time varying, and thus changing the maximum achievable data rate Rmax,i . In our work, we use the long-term channel quality to estimate the Rmax,i of each user i, for example, an average Signal-to-Noise Ratio (SNR) over 1 second period. As a result, any drastic change of channel condition, which may occur in a very short period of time, can be neglected, as the average channel quality is slowly changing. Figure 2 shows the relationship between the estimated Rmax and the average SNR. The data rate is varied by applying different modulation and coding schemes, where Rc is the coding rate of a convolution code. The estimated Rmax,i of each user i is transferred to the QoE module as

U = f (R), f : R → M OS

(2)

where R is a set of possible application data rates. Changes of video data rate can be done in many different ways, for example, by varying the quantization of the encoder at the server, through transcoding, or even packet dropping. In this paper, we use a simple decoding and re-encoding (transcoding) technique, as it causes minimal video quality degradation when compared to the packet dropping scheme [11]. Furthermore, in constrast to the source-encoding rate adaptation, transcoding provides faster response when changing of the data rate is required immediately to resolve the network congestion problem which may occur either in the core network or at the base station. At each video data rate, we measure the

13

4.5 4

U(MOS)

3.5 News MthrDotr Foreman Soccer

3 2.5 2 1.5 1 0

100

200 300 Data rate (kbps)

400

Fig. 3. Scatter plot and linear/non-linear regression of the VSSIM-based video quality assessment model on VQEG Phase I test dataset. [12]

Fig. 4. Video utility functions for different video sequences obtained with transcoding.

video quality using the Video Structural SIMilarity (VSSIM) index [12]. Transformation of the objective video quality (e.g. VSSIM) to the predicted user perception in quality degradation (DMOS) can be done in several ways. For example, Video Quality Expert Group (VQEG) in ITU recommended to use a nonlinear regression function as shown in Figure 3. DMOS 0 means that the user does not see the quality degradation compared to the perfect video quality. Whereas, the higher DMOS refers to the lower video quality that the user would rate. Since all test video sequences that are used in the VQEG Phase I test [13] are not available publicly, for simplicity, we map the VSSIM value to the MOS scale using a linear function with an upper and lower bound as follows: ⎧ 1, if V SSIM < 0.74 ⎪ ⎪ ⎨ M OS = a · V SSIM + b, if 0.74 ≤ V SSIM ≤ 0.98 ⎪ ⎪ ⎩4.5, if V SSIM > 0.98

then formulate the problem of tuning mechanism between the system efficiency and the user quality fairness. A. Utility maximization (MaxSum) Network operators are usually interested in their system performance, for example, how efficient their wireless resource is shared among users. In this paper, we measure the system efficiency by calculating the mean perceived quality of all users. In other words, the system efficiency is defined as the total sum of utility of all N users. The objective function of the optimization problem aiming at maximizing the system efficiency in a network resource constrained system can be described as follows: N  Ui (αi )) αopt = arg max( α

(4)

i=1

N subject to i=1 αi = 1 where αi ∈ [0;1] is the wireless resource share of user i, and α is the vector of all users’ resource share. αopt is the optimal resource allocation that maximizes the objective function. Ui (αi ) is the utility function of the given resource share αi of user i.

(3) Note that the upper and lower bound, and the constant parameters (a and b) of the linear function are determined such that it approximately fits best to the scatter plot of objective and subjective measurements as depicted in Figure 3. Figure 4 depicts examples of video utility curves for four different video sequences. In this example, we assume that the streaming server provides a high video quality with high data rate (e.g. ’News’ at 350kpbs, ’Soccer’ at 450kbps), and the data rate is adapted to lower rates by applying the video transcoding technique. All videos have QCIF resolution and are encoded with H.264 AVC at 30 frames/sec. We observe that the user perceived quality for the high motion video content (e.g. ’Football’ video) is more sensitive to the rate adaptation than the low motion video content (e.g. ’News’ video). This implies that difference in video content has a lot of impact to the user perceived quality and thus applying innetwork rate adaptation should take this impact into account.

B. Utility max-min fairness (MaxMin) With the utility max-min fairness, the network resources are distributed such that all users experience the same perceived quality regardless of the application type, content type and wireless channel condition. For a resource constrained system, the utility max-min optimization problem can be described as follows:  (5) αopt = arg max min Ui (αi ) subject to

N

i=1

α

i∈S

αi = 1

C. Efficiency vs Fairness trade-off problem formulation To formulate the tuning problem, we define the system efficiency and the quality unfairness as follows: Definition 1: The system efficiency e is the total sum of the MOS values perceived by all N users. Definition 2: The quality unfairness k is the maximum MOS deviation among users in a user group S, which is the

IV. Q O E- DRIVEN RESOURCE ALLOCATION OPTIMIZATION In the following subsections, we briefly discuss the two objective functions that have been used in our previous work [7]: the utility maximization and the utility max-min. We

14

Fig. 5. Schematic depiction of the proposed tuning mechanism between system efficiency e and quality unfairness k.

Fig. 6.

perceived quality difference between the user experiencing the highest MOS (M OSmax ) and another user experiencing the lowest MOS (M OSmin ) at each time instance. Figure 5 shows how the two objectives (utility maximization and utility max-min) can be mapped on a two-dimensional diagram capturing the system efficiency and the quality unfairness. In this example, the utility maximization results to the highest system efficiency and the lowest quality fairness. Whereas, the highest fairness can be achieved by using the utility maxmin, however, it comes at a cost of system efficiency. The operation points of maxmin and maxsum are denoted with index 1 and 2 respectively, which then specifies the fairness interval F and the system efficiency interval E. Alternative to achieving maximum system efficiency or maximum quality fairness, the network operator may prefer to have any intermediary operation point or to set a desired fairness level and a targeted system efficiency. Hence, an algorithm that enables a control of operation point is necessary. The desired operation point p is assumed to lie within the range of F and E.

Optimization progress with the Sum-MOS algorithm.

on the utility max-min resource distribution [4]. From this point, the algorithm improves the total sum of quality of all users by using the greedy-based utility maximization (MaxSum MOS) [7], and we stop the algorithm until the target mean MOS is achieved. Unless, if the target mean MOS is set too high, the algorithm will stop at the maximum sum of perceived quality of all users that is achievable. Figure 6 shows how the result is developed in each optimization step when applying the Sum-MOS algorithm with an example of setting the system efficiency requirement ereq to 102. In this scenario, we have 30 users accessing different video contents and experiencing different wireless channel condition at a time instance. B. k-algorithm As its name indicates, the k-algorithm focuses on the quality (un)fairness k. It allows the network operator to apply a strict fairness constraint kreq that is set in advance and to allocate its network resource accordingly, while maintaining the system efficiency e as high as possible under this fairness condition. In case, the algorithm cannot find an operating point that meets the fairness constraint, for example, due to the discrete application utility function, it will find the operation point closest to the desired fairness operation point. The k-algorithm is summarized in Algorithm 1, and depicted in Figure 7. The same scenario of 30 users as mentioned in Section V-A is applied here. For this optimization, we set the kreq to 1 as an example of the fairness constraint. From the results, we observe that there is a period, in which the k-constraint is not fulfilled. As the k-algorithm starts with the zero resource allocation, this can happen when the utility point for the low data rate is not provided. In this example, this is caused by the ’Foreman’ video as shown in Figure 4.

V. E FFICIENCY AND FAIRNESS TUNING MECHANISM Due to the complexity and complication that the tuning parameters (e.g. E, F ) are strongly dependent on the scenario specified by the number of users, channel variation among users, and application utility functions, we propose the tuning algorithm based on a heuristic, iterative solution. From the two criteria as discussed in section IV-C, we devise three tuning algorithms allowing the network operator to find a desired operation point of resource allocation by specifying A) only e-constraint, B) only k-constraint, or C) both e and k constraints at the same time. The search for optimal resource allocation that meets the given constraint(s) is determined by using the greedy search algorithm, which leads to close-tooptimal results [15]. We elaborate each of the tuning algorithm in the following subsections.

C. Advanced k-algorithm (e and k constraints) Advanced k-algorithm allows the network operator to allocate its network resource with pre-defined ereq and kreq constraints. Since it is possible that both constraints may not be met from any feasible set of resource allocation, we propose a two-step optimization that combines the Sum-MOS algorithm with the k-algorithm. We start with the k-algorithm as described in Algorithm 1 in order to fulfill the fairness criteria. From this point, we check whether the mean quality

A. Sum-MOS algorithm (e-constraint) The Sum-MOS algorithm enables a full control of resource allocation in order to deliver the desired mean quality (or sum of quality) of all users that is pre-defined by the network operator. To avoid large quality differences among the users, the Sum-MOS algorithm starts allocating the resources based

15

Algorithm 1 k-algorithm

2:

4: 6:

8: 10:

12: 14:

16: 18:

Input: Resource allocation of N users α = [α1 ,α2 ,...,αN ], step size of resource Δα, video utility function U , quality unfairness constraint kreq . Output: Optimal operating mode αopt . Initialization: zero resource share: α = [0, 0, ..., 0], M OS = [1, 1, ..., 1]. loop for i = 1 to N do Calculate the difference to next MOS level and the required resource: ΔM OSi , Δαi OSi Compute utility gain Gi = ΔM Δαi end for Ordering all users by utility gain G Find the user giving highest utility gain (nmax ) Precalculate Nthe fairness knew and the total allocated resources αnew = i=1 αi by assuming that resource share is given to nmax if knew ≤ kreq and αnew ≤ 1 then Allocate Δα to the user nmax else Continue searching for the user nnext with smaller G but satisfying the kreq and resource constraints end if break if there is no resource left end loop

Fig. 7.

of all users is equal or greater than the desired mean quality ereq . If this is the case, the resource is allocated as of the output of k-algorithm. Otherwise, based on the result of kalgorithm, we proceed with the 2nd-step optimization aiming at utility maximization (Max-Sum MOS [7]). The optimization stops either when the sum of perceived quality of all users reaches at the level of ereq or when the maximum sum of perceived quality of all users is achieved. The latter case will happen when the ereq is set too high. If the result of the 2ndstep optimization dissatisfies the kreq , we find a new operation point that is based on the priority of system efficiency ce and quality fairness ck . We assume that the priority is specified in advance by the network operator. Figure 8 illustrates how the advance k-algorithm searches for the desired operation point based on the ereq and kreq constraints from the same scenario with 30 users as mentioned in Section V-A. In this example, we set ereq and kreq to 103 and 1 respectively. Obviously, it is impossible to find an operation point that fulfills both constraints. By applying the priority values (e.g. ce = 0.7, ck = 0.3), the algorithm can search for an operation point that best fits to the compromise operation point based on the given priority.

Fig. 8.

VI. S IMULATION RESULTS As mentioned in Section I, we consider a resourceconstrained single LTE base station scenario, in which 20 users are watching a video on their terminals and are experiencing different wireless channel conditions. The MATLAB-based LTE simulator [8] is used to simulate the mobile system and to verify feasibility of the proposed tuning algorithm. All parameters used in our simulation are given in Table I. Figure 9 shows the mean MOS e and the unfairness k as a function of

16

Optimization progress with the k-algorithm.

Optimization progress with the advanced k-algorithm.

time, when applying the utility maximization (MaxSum), the utility max-min optimization (MaxMin), and the Sum-MOS algorithm with different settings of ereq constraints (e-algo). Obviously, with the MaxSum, the differences in perceived quality among all users are much larger than applying the MaxMin, however, the MaxSum has a higher mean MOS of all users in the system. With the e-algo, the target mean MOS ereq is maintained during the simulation period, for example at 3.3 mean MOS. In case, the ereq is set too high and cannot be achieved, it allocates resources so as to achieve a possible maximum mean MOS as result in the MaxSum case. In contrast, with the k-algorithm with different kreq , the system maintains the difference in quality among users over time as shown in Figure 10. However, this results to a variation of the mean MOS over time. We observe that the higher kreq we set, the higher mean MOS of the system we could achieve. Figure 11 shows how the mean MOS and the unfairness develop over time when applying the advanced kalgorithm with ereq and kreq set to 3.3 and 0.35 respectively. We see that when both requirements cannot be met at the same time (e.g. at 6th sec), the priority parameters ce and ck play an important role. For instance, if the ce is set to a higher priority, the resulting mean MOS is getting closer to the ereq and vice versa. From all simulation results, one can observe that the user fairness comes at a cost of the system efficiency. It should be also noted that all tuning mechanisms do not intend to get a better result than the MaxSum and the MaxMin in terms of the mean MOS and the user fairness respectively. In fact, they are used to provide the network operator a flexibility

TABLE I S IMULATION PARAMETERS Mean MOS

OFDMA full duplex FDD 2.8W 1.08 MHz 72 3GPP-LTE turbo downlink 35m-500m WINNER urban macro-cell ”C2” [14] 12 x 7 PGPS [9] {0, 30, ..., < 500}kbps H.264 Copy previous frame Transcoding

5

6

7

Mean MOS

Time (sec) MOSmax − MOSmin

(II)

MaxMin 8 MaxSum 9 10 e−algo: ereq=3.3 e−algo: ereq=3.4

2 1.5 1

5

6

7

2

MaxMin 8 9 10 MaxSum advk−algo: ce=0.7, ce=0.3 advk−algo: ck=0.3, ce=0.7

1.5 1 0.5 0

1

2

3

4

5

6

7

8

9

10

R EFERENCES

0

1

2

3

4

5

6

7

8

9

[1] H. Jiang, W. Zhuang and X. Shen, ”Cross-layer design for resource allocation in 3G wireless networks and beyond,” IEEE Comm. Magazine, vol. 43, no. 12, pp. 120-126, December 2005. [2] A. Eryilmaz and R. Srikant, ”Fair Resource Allocation in Wireless Networks Using Queue-length-based Scheduling and Congestion Control,” IEEE INFOCOM, Miami, FL, USA, Mar. 2005. [3] F. P. Kelly, ”Charging and rate control for elastic traffic,” European Transaction for Telecommunication, vol. 8, pp. 33-37, Jan. 1997. [4] Z. Cao, et al., ”Utility max-min: An application-oriented bandwidth allocation scheme,” IEEE INFOCOM, New York, NY, USA, Mar. 1999. [5] S. Khan, et al., ”MOS-based multiuser multiapplication cross-layer optimization for mobile multimedia communication,” Advances in Multimedia, article ID 94918, 2007. [6] A. Saul, ”Wireless resource allocation with perceived quality fairness,” IEEE Asilomar, PACIFIC GROVE, CA, USA, Nov. 2008. [7] S. Thakolsri, et al., ”QoE-Driven Cross-Layer Optimization for High Speed Downlink Packet Access,” Journal of Communications, Special Issue on Multimedia Communications, Networking and Applications, vol 4, no 9, pp. 669-680, Oct. 2009. [8] A. Saul, et al., ”Cross-Layer Optimization With Model-Based Parameter Exchange,” IEEE ICC, Glasgow, Scotland, UK, Jun. 2007. [9] A. K. Parekh and R. G. Gallager, ”A generalized processor sharing approach to flow control in integrated services networks-the single node case,” IEEE INFOCOM, Florence, Italy, May 1992. [10] ITU-T Recommendation P.800, ”Method for subjective determination of transmission quality,” Aug. 1996. [11] S. Thakolsri, W. Kellerer and E. Steinbach, ”QoE-based Rate Adaptation Scheme Selection for Resource-constrained Wireless Video Transmission,” ACM Multimedia, Florence, Italy, Oct. 2010. [12] Z. Wang, L. Lu and A. C. Bovik, Video Quality Assessment Based on Structural Distortion Measurement, Signal Processing: Image Communication, vol. 19, no. 1, pp. 121-132, Feb. 2004. [13] VQEG, Final report from the video quality experts group on the validation of objective models of video quality assessment, Mar. 2000. [14] IST-2003-507581 WINNER, ”D5.4 final report on link level and system level channel models,” ver. 1.4, Nov. 2005. [15] S. Khan, et al., ”QoE-based Cross-layer Optimization for Wireless Multiuser Systems,” In proc. 18th ITC Specialist Seminar on Quality of Experience, Karlskrona, Sweden, May 2008.

10

Fig. 9. Comparisons of the Sum-MOS algorithm and the MaxSum/MaxMin algorithms: (I) Mean MOS of all users over time, (II) Unfairness over time. (I) 3.5

Mean MOS

4

0.5

Time (sec)

3.4 3.3 3.2 MaxMin 3.1

1

2

3

4

5

6

7

Time (sec)

8 MaxSum 9 10 k−algo: kreq=0.35

(II)

MOSmax − MOSmin

3

them by introducing additional constraints of system efficiency and user fairness that are defined by the network operator prior to the optimization of network resource allocation. Three variants of tuning algorithm are introduced, which allows the network operator to set their requirement differently, e.g. only the system efficiency, or only the user fairness, or both of them. The proposed tuning algorithms are implemented in the simulation of an OFDMA system, and its feasibility is proved by the simulation results.

3.2 4

2

Fig. 11. Comparisons of the advanced k-algorithm (with ereq =3.3 and kreq =0.35) and the MaxSum/MaxMin algorithms: (I) Mean MOS of all users over time, (II) Unfairness over time.

3.3

3

1

Time (sec)

3.4

2

3.2

(II)

(I)

1

3.3

Time (sec)

3.5

3.1

3.4

3.1

MOSmax − MOSmin

Transmission scheme Duplexing Total transmit power Channel bandwidth Number of subcarriers Mapping and coding Distance from eNodeB Channel model Slot size (freq. x time) Scheduler Possible video rate Rvs Video codec used Loss concealment Video rate shaping

(I) 3.5

k−algo: kreq=0.65

2

k−algo: kreq=1.0

1.5 1 0.5 0

1

2

3

4

5

6

7

8

9

10

Time (sec)

Fig. 10. Comparisons of the k-algorithm and the MaxSum/MaxMin algorithms: (I) Mean MOS of all users over time, (II) Unfairness over time.

in specifying its policy for allocating its network resources in different ways based on the pre-defined ereq and kreq . VII. C ONCLUSION We introduce a novel tuning mechanism that allows the network operator to adapt its network resource allocation based on its policy in terms of the mean perceived quality of all users (system efficiency) and the quality fairness among users. The proposed algorithm makes use of the utility maximization and the utility max-min fairness, and extends

17