Service-oriented Cross-layer Management for Software ... - IEEE Xplore

4 downloads 949 Views 1MB Size Report
but also provide a software-controlled service management framework for cellular networks. I. INTRODUCTION. The traffic of mobile cellular networks is growing ...
2014 IEEE 25th International Symposium on Personal, Indoor and Mobile Radio Communications

Service-oriented Cross-layer Management for Software-defined Cellular Networks Xuan Zhou∗† , Zhifeng Zhao∗† , Rongpeng Li∗† , Yifan Zhou∗† , and Honggang Zhang∗†‡ ∗ York-Zhejiang

Lab for Cognitive Radio and Green Communications of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China ‡ Universit´e Europ´eenne de Bretagne & Sup´elec, Avenue de la Boulaie, CS 47601, 35576 Cesson-S´evign´e Cedex, France Email: {zhouxuan, zhaozf, lirongpeng, zhouyftt, honggangzhang}@zju.edu.cn † Dept.

Abstract—Rapid growing demand for mobile traffic and severe service bursts generated by various mobile applications are challenging the capacities and management of future cellular networks. The measurement data from real cellular networks indicate that the mobile Internet traffic expresses long-range dependence (LRD) characteristics, which differs from the traditional exponential assumption as well as 3GPP reports, would greatly deteriorate the experience of the subscribers. On the other hand, the mobile cellular networks treat most services without differentiation, although different services have their own requirements on transmission rates and delay. After introducing a new metric called the Quality of Experience (QoE) index for different services and analyzing the influence of LRD traffic, two approaches are proposed to alleviate the experience deterioration caused by traffic bursts: cloud based baseband resource pool to improve the flexibility of networks and wireless context-aware service controller to manage the distinctive QoE requirements of different services. Finally, simulations validate that these approaches not only contribute to service-oriented cross-layer optimization methods to satisfy the network resource constraints, but also provide a software-controlled service management framework for cellular networks.

I. I NTRODUCTION The traffic of mobile cellular networks is growing rapidly with the popularization of mobile devices, while operators are building more and more base stations (BS) with innovative technologies to provision the massive data. However, these dense BSs bring about strong interferences and expensive capital expenditures (CAPEX), and the fluctuation of traffic incurs the BSs’ underutilization. Besides, the increase of BSs’ capacity by operators lags far behind the growth of traffic demands, especially in dense metropolitan areas. Meanwhile, with the surge of over-the-top (OTT) services, the behaviors of using the mobile devices have been changed remarkably. Many mobile operators still rely on deploying more BSs to providing more capacity for huge data, without considering evolution of traffic models and variation of users’ demands. Therefore, by traditional planning method, traffic bursts are more severe than network designers expected, which may make the network congested even temporarily unavailable [1]. On the other hand, users would not complain if the instant messaging (IM) service is delayed with several seconds, however they would be unsatisfied if their payment on websites get stuck even if the background downloading is fast. Therefore, it is necessary for network designers to understand

978-1-4799-4912-0/14/$31.00 ©2014 IEEE

new behaviors of subscribers to provide better experiences in terms of service requirements. Recently, many operators have adopted policy control and charging (PCC) [2] architecture to manage the quality of service (QoS) using parameters such as: QoS class identifier (QCI) as well as allocation and retention priority (ARP). However, in practical cellular networks, OTT services are allocated to default bearer with non-guarantee bit rate (non-GBR) resource type, and the QCI for non-GBR services couldn’t describe the QoS requirements of numerous OTT applications. Moreover, due to the one-way control mechanism of PCC that the wireless feedback information is not even considered, networks determine the QoS policies only by static services’ configurations and subscribers’ profiles. In order to improve the performances of TCP and UDP in cellular networks, approaches like wireless-TCP and C-UDP are proposed [3][4]. However, these approaches are usually implemented in BSs as distributed solutions, which enhance end-to-end performance regardless of service discriminations; centralized solutions find difficult to collect enough wireless information owing to the BS output limitation and backhaul bottleneck. Moreover, both the TCP and UDP schemes on cellular networks don’t consider the scenario where radio resources are limited, and ignore the fact that the QoE of important services and subscribers need be guaranteed in priority. Conventionally, the software of BSs need to be updated for the services, if the priorities of specific services are raised. Besides, cellular operators are disinclined to select the BSs from various vendors, thus breaking the consistence of service management and the compatibleness of network elements. Cloud computing platforms have been adopted in various networks as infrastructures, meanwhile cellular networks are also migrating themselves from specific platforms to the general architectures. Baseband resource pool of cellular network on cloud computing platform makes mobile broadband capacity shared to wherever its needed, such as Cloud RAN from Huawei, Liquid Radio from NSN, Light Radio from AlcatelLucent, Cloud Radio from ZTE and C-RAN of China Mobile [5]. Recently, software-defined networks (SDN) solutions have been carried out in core networks to establish and manage new virtualized resources and networks without deploying new hardware technologies. Inspired by this benefit, SDN realizations in mobile cellular networks emerge as well [6], nevertheless with difficulties due to the complexity of air

2093

Since the networks deliver traffic data without service discrimination, some delay sensitive services may be influenced by insensitive services as well as unimportant services.

IM NegExp fit Logn fit Estimated WebBrowsing

−2

PDF

10

−4

10

Mean interval

−6

10

0

1

10

10

2

10 interval (Second)

3

4

10

10

Cell traffic (Mbps)

(b) Cell traffic variation. 60

800

40 20 0

50

100

150 200 250 300 t (Second) (c) Baseband resource pool traffic variation.

50

100

350

400

350

400

600 400 200 0

Fig. 1.

II. T RAFFIC M EASUREMENTS AND M ODELS Our dataset of measurements is based on 3 month anonymous traffic records collected from 7 million subscribers, originated from about 20000 GSM, UMTS and LTE BSs of China Mobile within a region of 3000 km2 . The dataset manifests that web browsing is the largest traffic contributing service consuming up to 1/3 of the total networks’ traffic, followed by IM (15.6%) and Social Networking Service (SNS) applications (6.8%). Besides that, the analysis results concerning the main attributes (e.g., inter-arrival time and message length) of the largest two services show that the traffic characteristics of real measurements were quite different from models proposed by 3GPP reports [7][8]. Specifically, Table I lists the main differences between our experimental results and the models in [7], and Fig. 1 (a) illustrates the inter-arrival time measurements of web browsing and IM of users and the corresponding fitting results. By fitting to the empirical data from the 2nd second to the 3000th second using the maximum likelihood estimation (MLE) method, the lognormal distribution is observed to best fit the inter-arrival time of IM service while the power-law (Pareto) distribution fits the inter-arrival time of web browsing service best. Notably, the peaks in the probability density function (PDF) of inter-arrival time of IM service are caused by keep-alive messages, which are periodically sent to servers to keep users timely notified once a message is received. This keep-alive mechanism is widely adopted by most IM and SNS designers nowadays, therefore, once the applications with keep-alive mechanism are prevalently installed on smartphones, a large percentage of BS resources would be occupied by the content-less messages.

(a) Inter−arrival time of services.

0

10

Pool traffic (Mbps)

interfaces of cellular networks. Fortunately, the cloud based baseband resource pool decouples the hardware and software of the BSs and provides real-time wireless information output conveniently, thus making cross-layer management of software-defined cellular networks (SDCN) feasible. In this paper, we firstly investigate the service characteristics based on real traffic measurement within a large-scale cellular network. Specifically, the heavy-tailed LRD characteristics of web browsing and IM traffic are verified, which is found to disaccord with the conventional models in 3GPP reports. Secondly, we propose a new objective named QoE index to describe the gap between the provided QoS and the required QoS with respect to different services, which could be utilized for service management. Our evaluations of the QoE indexes of subscribers further show that the conventional inaccurate service modeling would lead to traffic underestimation in network planning, which would lead to QoE deterioration. Thirdly, based on the new architecture for cross-layer management and the concept of SDN, we adjust parameters of transport layer in core networks to adapt to the QoS requirements of services constrained by network resources. Finally, benefiting from the real-time wireless information feedback of the baseband resource pool and QoE index models, our simulations show that cellular networks could achieve the service-oriented crosslayer management more effectively.

150

200 t (Second)

250

300

Inter-arrival time and traffic variations of services.

TABLE I D ISTRIBUTION MODEL DIFFERENCES . Parameter

3GPP Model

Measurement results

Browsing session inter-arrival time Browsing session length IM Message inter-arrival time IM Message length IM keep-alive period

Exponential Geometric Exponential Geometric Constants

Power-law Power-law Lognormal Power-law Discrete constants

Table I and Fig. 1 also provide an astonishing observation that the distributions of inter-arrival time of web browsing and IM are not exponential, but in the form of heavy-tailed distributions, which reflects the bursty characteristic in the time domain. In other words, users would generate many service demands in a short time yet be silent for a long time. Specifically, Fig. 1(b) and (c) depict the traffic volume over Iu interface of a NodeB and a RNC in busy time, whose peakto-average-ratio (PAR) is 4.15 and 1.51 respectively. Furthermore, most cellular traffic models such as [7] generally assume the inter-arrival time and service time to be independent with other sessions or messages. However, we find the traffic of IM service exhibits LRD attribute based on the real measurements [8]. Indeed, if a user has been busy/idle for a period of time, that user would possibly continue being busy/idle afterwards. The LRD characteristic also implies a social property that users may generate bursty traffic at the same time, such as the photo sharing messages when some events happen, and the user contacts when a train arrives. The traditional QoS describes the transmission performance of service delivery and does not take into account the users’

2094

Rij

M X X i=1

wj Qji /M

(3)

j

where weight wj indicates the importance of the QoE to the subscriber, and M is the number of subscribers within the cell. Apparently, the QoE index in LTE networks has much better performance than the other two modes, and the capacity of UMTS is about 100 subscribers larger than that of EDGE when the QoE indexes are set to be the same. In Fig. 3(a), a (a) The QoE index function of IM and web browsing service. (b) QoE indexes of all users within a cell Bit rate (Kbps) 25 0 100 200 300 6 EDGE UMTS LTE

20

(1)

f1IM(R)

4 QoE index

Qji = fj (Rij , dji )

the average of users’ QoE indexes in various networks as:

dji

where and are the instantaneous bit rate and time delay of the j-th service of the i-th subscriber, respectively. It is worthy to note that variables in Eq. (1) are time-variant, which means that delays and rate falls could be perceived simultaneously by QoE indexes deterioration.

f1WB(R) f2IM(d)

2

QoE index

actual sentiments to different services’ requirements. Comparatively, QoE is a subjective measure of a customer’s experiences from a service, based on psychology, cognitive science, engineering, etc., and thus is difficult to be directly evaluated merely by network measurements. In order to calculate the objective factor of QoE, we introduce a new QoE index model to describe the real-time gap between the provided QoS and the required QoS. Different services have their own requirements about the bit rate and time delay. For example, downloading service is more delay tolerable than web browsing, and IM service is less sensitive to the bit rate changes [9]. Hereinafter, the real-time QoE index of the service j to the i-th subscriber within the networks could be defined as:

15 10

f2WB(d) 5

0 0

10 20 Delay (second)

Fig. 2.

0 100

30

200 300 400 Users number

500

The QoE index function and QoE indexes value within a cell

III. I MPACT OF B URSTY TRAFFIC

Qji = f1j (Rij ) ∗ f2j (dji )

(2)

where f1 and f2 are defined as piecewise linear functions. Fig. 2(a) shows the corresponding variations of the QoE index with respect to f1 and f2 , where the QoE index could be from 0 (worst) to 5 (best). Intuitively, the QoE index of the service is proportional to its transmission rate, and is inversely proportional to its delay. Fig. 2(b) depicts our simulations of

400 seconds traffic sample in a NodeB during busy hours is illustrated, while traffic demands reach the ceiling of the BS capacity several times. The QoE indexes calculated by Eq. (3) during the 400 seconds are plotted in Fig. 3(b), which shows the QoE deterioration when the BS overloads. Besides, the QoE indexes decrease exceptionally from the 70th to 111th second when the BS traffic is relative light. Afterwards, we find that a website is temporarily unreachable during that time, which have impact on the QoE indexes of users within the cell as well.

Cell traffic (Mbps)

(a) A sample of BS traffic in busy hours. 60 40 20 0 50 100 150 200 250 300 350 t (Second) (b) QoE index variation corresponds to the traffic in Fig. 3(a)

400

30 QoE index

The influences of heavy-tailed LRD traffic, which typically include overly optimistic performance predictions and inadequate network resources allocation on wired networks, have been investigated by many researchers, and exert significant power on network management [11]. When we design and optimize cellular networks according to the traditional method, service models are assumed as in [7] and traffic volumes in peak hour are taken into account to determine the required maximum capacity. This estimation method is widely implemented and proven effective in circuit-switch (CS) call and early packet-switch (PS) services. Fig. 1 depicts that the mean inter-arrival time from our measurements is about 292 seconds marked by the vertical green line, and the corresponding exponential distribution estimated by the mean inter-arrival time is plotted as the red dot line. If we design the cellular network by assuming the inter-arrival time distributed as exponential, BSs would be overloaded by the underestimated burst traffic, particularly in the busy public areas. We calculate the QoE index value in Eq. (1) using our traffic model from the real measurements to describe the bursty traffic’s impacts when users aggregate. Referring to [9], for simplicity, we rewrite Eq. (1) as:

20

10

0

50

100

150

200 250 t (Second)

300

350

400

Fig. 3. A sample and corresponding QoE index of BS traffic in busy hours.

IV. A RCHITECTURE FOR C ROSS - LAYER M ANAGEMENT If the networks were designed to provide services to all the burst requests, we need to deploy much more BSs. However, our real traffic measurements show that even the busiest BS keeps at full load for only 2 hours, and 25% BSs are near

2095

complete idle during the whole day. In another real WCDMA traffic measurement in a major European city, 90% cell congestion durations are shorter than 1.2 seconds, and 90% cell congestion separations are below 780 seconds [10]. Therefore, given the limited operating expense (OPEX) and marginal effect of capacity expansion, operators would be considerate to enhance network capability for burst demands. Moreover, the transmission rate and delay control of insensitive services would be effective for alleviating the short bursts. We consider two approaches to alleviate the deterioration by bursts: cloud based baseband resource pool to improve the flexibility of networks and wireless context-aware service controller inspired by SDN concept to manage the distinctive QoE requirements of different services. In view of unbalanced spatial traffic distributions and short time spans of bursts, management of the baseband processing units (BBUs) of up to tens of cells could be centralized, connecting with the distributed remote radio heads (RRHs)[5]. The upper part of Fig. 4 depicts the architecture of the baseband resource pool, which consists of general-purpose processor platform, real-time operating system and baseband software. After the distributed BBUs migrate to the centralized baseband resource pool, the network traffic curve would change from Fig. 1(b) to Fig. 1(c) being more smooth. Moreover, the BS sizes could be much smaller and the deployment becomes easier. Besides, coordination mechanisms such as coordinated multipoint transmission (CoMP) depending on the baseband resource pool may enhance the throughput better than traditional solutions, since BSs are able to share information on the same platform more conveniently than via the external X2 interfaces. As we know, the conventional cellular networks conceal wireless details to applications by layers such as packet data convergence protocol (PDCP) layer, thus making the transportation layer difficult to detect the reachability of the peer end accurately and timely. The most benefit of the baseband pool is that the real-time wireless and resource information output on BS level and user level becomes possible, which is quite important for us to manage the QoE of the services. Take LTE network for example, a portion of the wireless information generated by the baseband resource pool is listed in Table II, which could be used for QoE management element (QME) to adjust parameters of transport layer reasonably. The quality type information in Table II indicates the quality of wireless links between UEs and BSs, which provides reference to the maximal transmission rate of the links. Information of the rest types in Table II forms the wireless context of UEs, which could be used for predicting quality of wireless links and user behaviors. The utilization information represents the timely status of BS and radio resources, which is important for service management. V. A LGORITHMS FOR Q O E ENHANCEMENT IN SDCN With the real-time measurement, configuration and signaling types information from the baseband resource pool, the QME could get the probability of failure pf of radio link control (RLC) and medium access control (MAC) layers, as well as

:LUHOHVVLQIRUPDWLRQ 5HDOWLPHRSHUDWLQJV\VWHP EDVHEDQGVRIWZDUH *HQHUDOSXUSRVH SURFHVVRUSODWIRUP

:LUHOHVVVWDWXVLQIRUPDWLRQ

40(

6HUYLFHIORZ SDUDPHWHUV

6XEVFULEHUFRQWH[W

6 /7(%DVHEDQG3RRO

6 00(

6*:3*:

*DWHZD\ 3&5)

875$1*(5$1%DVHEDQG 3RRO

*E

*Q 6*61

**61

Fig. 4. Architecture of the heterogeneous cellular network to realize QoE index management. Upper: cloud based baseband resource pool with wireless information output. Lower: interfaces and information flows of the network. TABLE II E XAMPLES OF WIRELESS INFORMATION OUTPUT. Type Quality Quality Quality Measurement Measurement Configuration Configuration Signaling Signaling Utilization Utilization Utilization

Wireless information

Channel

Channel Quality Indication Sounding Reference Signal HARQ Information Reference Signal Receive Power Timing advance and AoA Master Information Blocks System Information Block RRC Connection Setup UE Capability Information Available UL/DL PRB number MCS of Transport Blocks Buffer Capacity

PUCCH/PUSCH PUCCH PHICH PUCCH PUCCH BCCH PDSCH PDSCH PUSCH N/A N/A N/A

the predicted duration of failure tf , which would be conductive to QoS management. Similar to the QoE index, the probability and duration of wireless transmission failure functions are time-variant, which could be written as: (pf , tf ) = g(rp , hf , ar , sig)

(4)

where rp is the reference signal receive power, hf is the failure rate of HARQ , ar is the retransmission rate of ARQ in acknowledge mode. In addition, referenced signaling indication sig is considered in Eq. (4) to provide more prior knowledge to the QME, such as: handover request and bearer activation. By

2096

the model of probability and duration of wireless transmission failure in Eq. (4), the bounds of rate and delay of ith subscriber could be found according to pf and tf . For simplicity, we use exponential and linear functions to restrict the values of rate and delay in Eq. (2) respectively [13]: Ri < αe−pf β , f or pf > pth di > γtf + δ, f or tf > tth

(5)

where pth and tth are the thresholds that trigger the deterioration of rate and delay respectively, and α, β, γ and δ are parameters of the failure-throughput model. Afterwards, according to Eq. (5), we need to find the best service bit rate and time delay for all the services of all users: M X X max( wj Qji ) i=1

s.t.

X

Rij

j

= min(Ci , Ai )

j M X X i=1

(6) Rijk = CBk

j M X

j Rij = CSP

i=1

where Ci is the real-time transmission capability of the subscriber i which is mainly decided by its channel quality, Ai is the UE aggregated maximum bit rate (UE-AMBR) of the i-th subscriber stored in home subscriber server (HSS) j [2], and CSP is the bandwidth of the service provider j in core networks. When coordination technology such as CoMP is used, data streams of the service may come from several BSs. Accordingly, Rijk indicates the transmission bit rate of the jth service between the ith subscriber and the kth BS, and CBk indicates the maximum transmission capacity for all subscribers within the kth BS. By real-time quality and utilization types information of the baseband resource pool, instantaneous optimal rate and delay could be determined by QME using Eq. (6). If the number of the BSs and service categories are less than 5, Eq. (6) would be easy to solve. The solution of Eq. (6) makes it possible to manage the services, depending on the transport layer. Notably, TCP parameters have been tuned to perform well in wired networks to control the transmission bit rate and delay, when packet losses occur mostly because of congestion. However, cellular networks also suffer seriously from delay due to fast fading channel and handoffs, while TCP mechanism responds to all exceptions by just invoking congestion control and avoidance algorithms, resulting in end-to-end performance degradation. Many researchers have discussed various solutions such as wireless-TCP, but still didn’t take into account the viewpoints of wireless information and QoE management [3]. Meanwhile, UDP transmission over wireless link is also optimized by researchers, whereas without considering constrains of network resources either [4]. As Fig. 4 illustrates, the QME is

implemented between the baseband resource pool and the gateway with deep packet inspection (DPI) function to recognize and control the service parameters such as bit rate and time delay, in terms of real-time information output from that pool. Specifically, receive window (RWND), congestion window (CWND), retransmission timeout (RTO) of TCP and forward delay of TCP and UDP in the gateway could be dynamically set to meet the global QoE index optimization goal as in Eq. (6). In cellular networks between UEs and gateways, congestions seldom occur because of the QoS guarantee for the private subnet, so CWND should be used as transmission acceleration rather than congestion control. For example, the initial CWND could be set larger so long as the wireless link and UE are normal. Hence the data could be sent in fewer rounds. Delay-insensitive services such as IM and SNS could be postponed for a while and other services such as web browsing are served in priority when the BS is of full load. The retransmission of TCP could be co-designed with the Hybrid Automatic Repeat Request (HARQ) status, so that the retransmission could be more effective due to the low layer exposure. In another case, the congestion on uplink may impede the TCP-ACK message of downward services, thus we could perceive this phenomenon from the baseband resource pool and adjust the RWND via QME to reserve resources for the downward services. VI. S IMULATIONS AND A NALYSES The QME is designed with two main functions: enhancing the transmission efficiency when networks are light loaded and providing the best QoE index to minimize the impacts of congestion when BSs are overloaded. Fig. 5 illustrates how the QME works when traffic loads of BSs, subscribers and service platforms (SP) vary. When the networks are light loaded, transmission could be accelerated by increasing CWND, which not only provides better QoE but also avoid the resources competition caused by service burst lately. However, once BSs are overloaded, the QME adjusts the parameters of the transport layer in the gateway to enhance the QoE index as in Eq. (3). Furthermore, artificial intelligence and machine learning can be implemented in the QME, such as traffic prediction and user behavior (e.g., demands, mobility, location, etc.) predictions. Since the duration of traffic bursts is short and the occurrence of bursts is frequent, the baseband resource pool should keep prompt response to the traffic variations. Based on channel models and throughputs described in [14], real constitutions of services from measurements and queue models with different service rate are employed to evaluate the effect of QoE index enhancement. As a representative case as in Fig. 6(a), the QoE indexes after optimization by QME are plotted in red curves, and Fig. 6(b) and Fig. 6(c) illustrate the QoE indexes of IM service and web browsing service respectively. We find that the QoE indexes of IM service has been decreased to release resources to the web browsing service, because the IM service is less sensitive to transmission performance, and the weight of IM service in Eq. (3) is smaller than which of the web browsing service. The

2097

&L&%N 5LMN&M63

(a) The global QoE enhancement after optimization. 25 EDGE UMTS LTE EDGEO

20

8SOLQNKHDY\

QoE index

/LJKW

10

'HFUHDVH5:1'

8(OLQNVWDWXV 8QVWDEOH

8(OLQNVWDWXV

2.

WI•WWK 5HVHW&:1'

WIWWK

8(ORDG

+ROG&:1' ,QFUHDVH572

/LJKW

LTEO

0 100

200 300 400 500 1000 Users number (b) The QoE index of IM service after optimization. (c) The QoE index of web browsing service after optimization. 25 25 20

+HDY\ QoE index

WI

UMTSO

5

8QVWDEOH

2.

15

&DFKH

20 QoE index

%6ORDG

+HDY\

15 10 5

,QFUHDVH&:1'

0 100

4R(LQGH[RSWLPL]DWLRQEDQGZLGWKDOORFDWLRQGHOD\H[HFXWLRQ

10 5

200

300 400 Users number

Fig. 6. Fig. 5.

15

500

1000

0 100

200

300 400 Users number

500

1000

QoE enhancement after optimizations.

The flowchart of QME procedure.

optimized QoE indexes in EDGE and UMTS networks are about twice more than those before optimization if the users number is greater than 300. However, the QoE indexes in LTE networks increases only by 5% even the number of users is 500, while reasons for the differences of QoE improvement in different generations of networks are as follows: firstly, the LTE networks have better ways to provide stable throughput for the network layer in fading channel environments, therefore the wireless information feedback is less important to QME in LTE networks than which in EDGE and UMTS networks. Secondly, the overheads of data channel allocation are negligibly small due to the channel sharing mechanism of LTE networks. As a result, the QoE gain by the delay and rate control of IM service in LTE is less than that in EDGE and UMTS. Thirdly, the effect of QME is more evident when the QoE indexes are low, however, the QoE indexes in LTE are not equally worse as the previous generations even when there are 500 users in a cell. Moreover, the enhancement by QME in LTE networks is more effective if the QoE indexes are less than 5. Compared with schemes of no wireless information feedback, our solution controls the service flows more accurately and achieves higher channel utilizations. VII. C ONCLUSION With the development of mobile networks, traffic bursts will be the key challenge of next generation cellular networks. The conventional service management demonstrates poor performance when numerous services with diverse requirements emerge. Therefore, inspired by cloud computing and SDN concepts, we envision a service-oriented management method on baseband resource pool to conduct unified optimization across the service layer, transportation layer, network layer as well as wireless layers (RLC, PDCP, MAC and PHY layers). In the paradigm of our work, the cloud based softwaredefined cellular networks are granted the capability to make service-oriented policies according to the service-specific endto-end QoE requirements and user-level wireless information.

Besides, our approaches provide a unified solution to manage both the QoE of subscribers and the QoS of networks. ACKNOWLEDGMENT This paper is supported by the National Basic Research Program of China (973Green, No. 2012CB316000), the Key (Key grant) Project of Chinese Ministry of Education (No. 313053), the Key Technologies R&D Program of China (No. 2012BAH75F01), and the grant of Investing for the Future program of France ANR to the CominLabs excellence laboratory (ANR-10-LABX-07-01). R EFERENCES [1] M. Donegan, “Android signaling storm rises in Japan”, Jan. 2012. [Online]. Available: http://www.lightreading.com/blog/mobileoperating-systems/android-signaling-storm-rises-in-japan/240005553 [2] 3GPP TR 23.203 V11.8.0, “Policy and charging control architecture”. [3] Y. Tian, K. Xu, N. Ansari, “TCP in wireless environments: problems and solutions,” IEEE Communications Magazine, vol. 43, no. 3, pp. 27-32, Mar. 2005. [4] H. Zheng, J. Boyce, “An improved UDP protocol for video transmission over internet-to-wireless networks,” IEEE Transactions on Multimedia, vol. 3, no. 3, pp. 356-365, Sep. 2001. [5] Chih-Lin I, C. Rowell, S. Han, Z. Xu, G. Li, and Z. Pan, “Toward green and soft: a 5G perspective,” IEEE Communications Magazine, vol. 52, no. 2, pp. 66-73, Feb. 2014. [6] A. Gudipati, D. Perry, L. Li, and S. Katti, “SoftRAN: Software defined radio access network,” in Proceedings of ACM HotSDN’13, Aug. 2013. [7] 3GPP TR 43.802 V0.4.0, “GERAN Study on Mobile Data Applications”. [8] Xuan Zhou, Zhifeng Zhao, Rongpeng Li, Yifan Zhou, Jacques Palicot, and Honggang Zhang, “Understanding the nature of social mobile instant messaging in cellular networks,” IEEE Communications Letters, vol. 18, no. 3, pp. 389-392, Mar. 2014. [9] ITU-T G.1010, “End-user multimedia QoS categories,” 2001. [10] 3GPP TR 22.805 V12.1.0, “Feasibility study on user plane congestion management”. [11] A. Erramilli, O. Narayan, and W.Willinger, “Experimental queuing analysis with long-range dependent packet traffic,” IEEE/ACM Transactions on Networking, vol. 4, no.2, pp. 209-223, Apr. 1996. [12] Rongpeng Li, Zhifeng Zhao, Xuan Zhou, Jacques Palicot, and Honggang Zhang, “The Prediction Analysis of Cellular Radio Access Network Traffic: From Entropy Theory to Networking Practice,” IEEE Communications Magazine, vol. 52, no. 6, pp. 234-240, Jun. 2014. [13] J. Padhye, V. Firoiu, D. Towsley, J. Kurose, “Modeling TCP Throughput: A Simple Model and its Empirical Validation,” ACM SIGCOMM Computer Communication Review, vol. 28, no. 4, 1998. [14] 3GPP TR 25.996 “Spatial channel model for MIMO simulations”.

2098

Suggest Documents