Pricing for QoS-Based Wireless Data Services and its Impact on Radio Resource Management Patrick Hosein Huawei Technologies Co., Ltd. 10180 Telesis Court, San Diego, CA 92121, USA
[email protected]
Abstract—The use of cellular networks for both voice and data communications is growing at a rapid pace throughout the world. Such growth was initially fueled by the need for mobile voice communication. However, with the advent of 2G networks and now with 3G networks, data services are quickly outpacing that of voice. Normally such growth is welcomed since it results in increased revenue for the wireless operator. However, unlike wireline networks for which capacity can easily be increased, the capacity of wireless networks is limited. The load that can be sustained by a wireless network depends on various factors but the major ones are cell size, bandwidth and spectral efficiency. Each of these have limits and so as the demand for data services increases, the pricing and allocation of the limited available resources become increasingly more important. In this paper we address the impact that limited resources have on pricing and how, in turn, radio resource management algorithms are affected. The emphasis in this work is on the overall framework aspects and not on the designs of the individual radio resource management algorithms.
I. I NTRODUCTION The demand for cellular services (e.g., LTE [1] and WiMAX [2]) is rapidly growing especially in developing countries. In these countries voice services are more economically provided over wireless rather than wireline networks. Prices for both handsets and infrastructure equipment has also significantly dropped making such services affordable. In developed countries the demand for data services continues to rise and the use of applications such as YouTube is creating greater demands on radio resources. Although the growth of human sources will eventually converge (finite population), the per user data demand as well as access by non-human sources (Machine to Machine) will both continue to grow. This projected increase in data traffic is investigated in a Cisco whitepaper [3]. Using data from that source we plot, in Figure 1, the forecasted data traffic (in Petabytes per month) as a function of time (in years) for various types of applications. We note that video traffic is expected to grow quite rapidly and will dominate all other forms of traffic. This is of concern because video traffic is resource intensive (high bit rate with delay constraints). Next in importance is web-based traffic followed by traffic that is delay sensitive but not very throughput intensive (P2P, gaming and VoIP). It is therefore important to design radio resource management algorithms with this expected change in traffic mix. The exponential growth in data traffic cannot be completely accommodated by increased spectral efficiencies (bounded by
the Shannon limit), cell size (bounded by hardware cost) or number of antennas (bounded by physical limitations). Capacity can be increased by using more bandwidth but this too is limited by the available spectrum. The whitepaper by Rysavy [4] investigates the growth in capacity as additional spectrum is made available for wireless services. Figure 2, taken from [4], contains a plot of total data demand and total available capacity as a function of time. This is the capacity that can be achieved as additional spectrum is made available to 4G networks. We find that at some point, the demand will exceed the capacity. Beyond this point resources must be appropriately managed and priced. The conclusion that pricing must eventually be used to better allocate limited resources is not new and in [4] the authors conclude that “Pricing plans are the easiest way for operators to limit data consumption. Plans that are too restrictive, however, and that prevent users from doing much of what they do over wireline connections will significantly constrain market development and the data revenue opportunity.” In that paper it is also suggested that offloading traffic to other technologies such as Femto cells can be helpful. Another suggestion is that, by optimizing frequently used applications (such as email), the over-the-air traffic can also be reduced. In this paper we investigate how pricing can be used to ensure that users continue to receive acceptable performance while allowing wireless operators to make a profit even when capacity becomes limited. In addition, we investigate how pricing plans must be modified to take into account QoS services that will be introduced in 4G networks. Finally, the impacts on radio resource management algorithms will also be investigated. In the next section we describe how QoS services are defined in the LTE standard [5]. The introduction of such a framework makes it possible to differentiate the quality of service achieved by different classes of users. Next we investigate suitable pricing models as well as the impacts these models have on radio resource management. II. U SER Q O S AS DEFINED FOR LTE The LTE wireless standard includes a framework for the provision of QoS services. In this section we briefly describe this framework since it will determine what pricing strategies are possible. A QoS Class Identifier (QCI) is a scalar that is used as a reference to a specific forwarding behaviour
2000
QCI
Video Web/Data/Other P2P Gaming VoIP
Petabytes Per Month
1500
Type
Priority
PDB
PLR
Example Conversational Voice
1
GBR
2
100ms
10−2
2
GBR
4
150ms
10−3
Live Streaming Gaming
3
GBR
3
50ms
10−3
4
GBR
5
300ms
10−6
Buffered Streaming IMS Signalling
5
Non-GBR
1
100ms
10−6
6
Non-GBR
6
300ms
10−6
TCP-based Voice, Gaming
1000
500
7
Non-GBR
7
100ms
10−3
8
Non-GBR
8
300ms
10−6
TCP-based
300ms
10−6
TCP-based
9
0 2009
Non-GBR
9
TABLE I S TANDARDIZED QCI S PECIFICATIONS 2010
2011
2012
2013
2014
Year
Fig. 1.
Global Mobile Data Traffic 2009-2014 by Application
Rate MBR
Fig. 2.
Projected Growth of Mobile Data Demand and Capacity
(resource type (Guaranteed Bit Rate (GBR) or non-GBR), priority, packet delay budget (PDB) and packet loss rate (PLR)) to be provided to a Service Data Flow (SDF). According to [5], “Services using a GBR QCI and sending at a rate smaller than or equal to GBR can in general assume that congestion related drops will not occur and 98% of the packets shall not experience a delay exceeding the QCI’s PDB. Services using a Non-GBR QCI should be prepared to experience congestion related packet drops, and 98% of the packets that have not been dropped due to congestion should not experience a delay exceeding the QCI’s PDB”. In certain situations (e.g., transient link outages) congestion related drops may still occur even for services using a GBR QCI with a rate no greater than GBR. The priority levels are used to prioritize servicing of flows during periods of congestion. The PDB is the maximum delay with a confidence level of 98%. The PLR defines an upper bound on non-congestion related packet losses. In the LTE standard a recommended list of QCI values is also specified. These are provided in Table I together with typical usage scenarios. In addition to the GBR value, a flow may also be assigned a Maximum Bit Rate (MBR). This rate is a strict limit placed on a flow. Therefore even if resources are available, once the
PDB
PLR PDB
dropped
Heavy System Load
Fig. 3.
PLR Moderate System Load
PLR PDB
PLR PDB Light System Load
GBR
dropped
Time
Illustration of PDB and PLR constraints for a GBR Flow
arrival rate of the flow exceeds this value, sufficient packets are dropped to maintain the carried rate below the MBR value. Given these definitions, various actions can be taken for different loading conditions. For example, if a flow exceeds its GBR value but resources are available then the scheduler may continue to maintain its PDB and PLR constraints. However, if resources are limited then the scheduler may instead queue packets but not drop them and hence attempt to maintain the PLR constraint but not the PDB constraint. Under even heavier loading conditions, once the GBR value is exceeded then packets can be dropped to maintain the throughput below GBR at the expense of an increased PLR. In this case the delays of packets that are delivered may continue to satisfy the PDB constraint. These scenarios are illustrated in Figure 3 where the rate achieved by the flow is plotted as a function of time. At various points in time we illustrate where packets are dropped and where the PDB and PLR constraints are satisfied (indicated with a tick) and where they may be violated (indicated with a cross). For a given flow, the operator must ensure that sufficient resources are provided to the flow so that the corresponding QoS constraints are satisfied. In this paper we assume that algorithms (scheduling, power control, interference management etc.) have been implemented to achieve this goal for each flow. Note that in order to maximize user capacity (and hence revenue), each flow should be provided with exactly the amount of resources needed to achieve its QoS constraints
pricing
Wireless Operator
revenue
in−network savings
User Population and Usage
load characteristics capacity engineering
satisfaction (churn)
utilization Network
Fig. 4.
Pricing and User Behaviour Dynamics
(no more and no less) so that the maximum number of users can be accommodated. In addition, congestion and admission controls must be developed to ensure that during busy hour periods QoS constraints can be maintained. Because of the bursty nature of packet arrivals as well as the fluctuation of the channel quality and interference experienced by each user’s transmission, sectors will typically run at subunity utilization in order to ensure that QoS constraints are met. The utilization would depend on the nature of the traffic. If there are many low rate flows then significant statistical multiplexing gains can be achieved and hence the system can operate at a high utilization. As more high rate flows are accepted, the statistical multiplexing gain decreases and more resources must be reserved for potential fluctuations. Pricing for a flow therefore also depends on the potential statistical multiplexing gain that can be achieved. III. P RICING AND U SER B EHAVIOUR One needs to take into account several factors when designing a pricing strategy. If pricing is too high then customers are not attracted and if too low then too many customers are attracted leading to capacity issues. Therefore pricing must be such that the operator can fully utilize its resources while providing acceptable performance to each user. The system dynamics is summarised in Figure 4. The user population and type of applications used determine the revenue achieved by the operator and the characteristics of the load placed on the network. It also determines user savings achieved by innetwork communications. The network must be operated at high utilization (for maximum revenue) but must also provide high user satisfaction to reduce churn (i.e, the fraction of users that switch to other networks because of poor service). The wireless operator must balance these objectives through appropriate pricing and network provisioning. Note that the system can easily be perturbed from a steady state point by various external factors (change in pricing of competitors, introduction of new services, introduction of new technologies etc.). With each perturbation the operator must be able to converge to a near-optimal steady state point. User behaviour can severely affect the convergence point of this dynamic system. For example, one popular pricing
plan that is presently offered is flat rate, unlimited volume service (but with no QoS). Earlier phones were not capable of consuming large amounts of data and hence such a plan was at one time suitable. However, with the advent of smart phones on which one can view videos or listen to radio broadcasts such an assumption no longer holds. Furthermore, such phones can be used to tether one’s laptop to the Internet and essentially use the wireless network in the same fashion as one would use an Internet connection at home. Since unlimited volume pricing plans are not sustainable then, at some point, volume limits must be placed and usage beyond this limit would be priced at a higher rate. This is already done by some operators. With data volume limits, users behave differently. As the end of the billing cycle approaches, if the volume is below the limit then the user will increase usage but if she exceeds the limit then usage will be reduced. If at the end of the cycle the volume far exceeds the limit then the user will consider switching to a higher volume limit plan. The operator’s objective will be to “encourage” the user to exceed its limit in order to increase revenue above the flat rate paid. Note that this type of behaviour is already exhibited with voice plans that have limits on monthly voice minutes. IV. R EVENUE M AXIMIZATION Given the standardized framework for QoS services and the effect of user behaviour on the system dynamics we next investigate a suitable pricing strategy. Present wireless network standards do not have support for QoS services. Best effort service is provided and flat rate pricing is used. One can obtain a flat rate service with unlimited data volume or with limited data volume (e.g., 10 Gbytes). If the user exceeds its allotted volume then additional traffic is charged a higher per bit rate. Therefore it is left up to the user to determine what package is most suitable. If the chosen volume limit is tool large then the user ends up paying for capacity that is not used while if the volume limit is too small then the user ends up paying a high rate whenever the volume limit is exceeded. Pricing plans should be as simple as possible for consumers. Therefore, we propose that pricing be made a (simple) function of both guaranteed performance (rate, delay, PLR etc.) and volume. The cost of serving a customer is proportional to the resources that must be reserved for use by the customer. For a GBR flow, the amount of reserved resources does not have to be that required to always maintain a rate of GBR. This is because the offered load will vary and so one can achieve some statistical multiplexing gains since not all flows will peak at the same time. However this multiplexing gain will decrease as the GBR value increases (since fewer flows would then be admitted) and so the utilization will decrease as the GBR values increase. This in turn implies that the cost per resource must also be increased in order to maintain the same total revenue. The net result is that the cost of a flow should grow super-linearly with the size of the flow. On the other hand, one typically provides discounted pricing for high GBR subscriptions which would mean sub-linear growth of
Cost
Gold (4 GB limit)
Silver (4 GB limit) Silver (2 GB limit) Volume Used (GB) 2 increase usage
Fig. 5.
4 decrease usage
change plan
Cost versus Usage for various Subscription Plans
price versus rate. Hence a reasonable approximation would be simply to assume that price grows linearly with rate. One can use a similar argument for cost as a function of monthly volume and so we also assume that cost should be a linear function of the volume limit of the subscription. We propose a small number of subscription plans. For each plan, QCI values are specified for all typical applications (video streaming, web browsing, VoIP, etc.). The cost per MB per month is specified for each of these plans depending on the QoS provided. A consumer is then free to pick a plan together with a volume limit and the monthly rate would be the plan rate times the data volume limit. If the volume limit is reached before the end of the billing cycle then a (higher) per bit rate is charged for the rest of the cycle. We illustrate this with the example depicted in Figure 5. Here we show the cost as a function of the volume used by the end of the billing cycle for three different users. Two are Silver users but with different volume limits and the third is a Gold user. In this case the per bit cost for excess traffic is the same for all three users but this does not necessarily have to be the case. Beneath the graph we indicate the actions that a Silver 2 GB limit user would take for different volume ranges. V. I MPACT ON R ADIO R ESOURCE M ANAGEMENT In the previous sections we outlined the overall price/user/operator dynamics that one can expect for future wireless networks. In this section we investigate how these affect the design of the various radio resource management algorithms. Such algorithms should provide the wireless operator with the flexibility needed to maximize its revenue while providing satisfactory service. A. Scheduler As described above, each user will subscribe to a pricing plan which will determine the QCI to be assigned to each flow of the user. For a given flow, the PLR depends on losses due to failed transmissions as well as losses due to dropped packets. Transmission failure rates depend on the mapping from reported channel quality information (CQI) to the modulation and coding scheme (MCS) used for transmissions to/from the user. It also depends on the number of retransmissions allowed but this value is typically determined by the PDB of the user (i.e., lower delay thresholds require fewer retransmissions).
Therefore the CQI to MCS mapping must be chosen to achieve the desired residual block level error rate (BLER). Packets are either dropped when the flow exceeds the GBR value or when resources are exhausted (congestion state). Therefore even if the flow does not exceed its GBR value, its queue may build leading to the dropping of packets that exceed the delay budget or exceed the buffer limit. The system must be provisioned and engineered so that the probability of congestion is sufficiently small so that the flow’s PDB and PLR can be maintained as per its QCI description. In the non-congested state, excess resources may be available and these can be assigned to flows to achieve performance above that which is guaranteed. In this case a decision must be made as to which flows should be allowed to exceed the GBR and by how much. We can answer this question given the previously described models. Assume that, as depicted in Figure 5, additional revenue is collected when users exceed their volume limits. The operator can then potentially increase revenue by serving a user’s flows as much as possible so that the user eventually exceeds her volume limit. We suggest a simple strategy whereby excess resources are allocated based on user channel conditions. In this way data consumption is maximized over these excess resources. This is typically referred to as maximum C/I scheduling. One can achieve all objectives (user satisfaction and revenue maximization) by using a utility-based scheduling framework (e.g., [6]). We provide a simple example of such a utility function but do not go into details of such a design. Consider a flow with a PDB of dmax , an average throughput of r and a spectral efficiency of µ. The corresponding flow priority (the gradient of the flow’s utility with respect to the resources allocated to the flow [7]) is given by P =
1 µ r (dmax − d)α
where d is the present estimate of the delay experienced by an arriving packet and α is determined solely by the PLR of the flow. Note that d is a function of r. If we let R denote the GBR of the flow and, for simplicity, assume a M/M/1 queuing system with service rate R and arrival rate r then we have d=
1 R−r
in which case we can write P =
µd . (Rd − 1)(dmax − d)α
Let us illustrate with the help of a simple example. We use a baseline case of µ = 1 bps/Hz, dmax = 150 ms, R = 100 kbps, and α = 0.5 (which we assume is the value that achieves the desired PLR). In Figure 6 we plot the priority of the flow as a function of the expected total delay of an arriving packet. We plot the baseline case as well as illustrate what happens as we vary various performance parameters of the flow. As the spectral efficiency increases the priority increases which happens to achieve our goal that under light loading (i.e., low delay values), scheduling is based primarily on spectral
5
Baseline Increased Spectral Efficiency Increased GBR Increased PDB Increased PLR
Scheduling Priority
4
3
2
1
0 20
Fig. 6.
40
60
80 Delay (ms)
100
120
140
Illustration of Priority Functions to achieve QoS
efficiency. As the GBR is increased the priority decreases because more resources are made available to the flow. As the PDB is increased the priority is again decreased since the flow can be allowed to build a larger queue. Finally as the PLR is increased the priority is decreased because the queue delay budget can be violated with a higher probability. Conclusion: Present schedulers (based on full buffer assumptions with an objective of Proportionally Fair rates) are unsuitable for future networks and flexible, delay dependent, utility-based schedulers should instead be investigated. B. Power Control and Interference Management Note that power control and interference management are closely related. The power used for each transmission (downlink or uplink) determines the success rate of the transmission but it also determines the amount of interference caused to neighbouring cells. The amount of interference caused by neighbours determines how much power is needed for a transmission to achieve a desired SINR. Furthermore, the variation of this interference affects the accuracy of the CQI used in determining the MCS for the transmission. Let us first consider interference caused within a site. For uplink transmissions, the basestation knows which users have transmitted as well as the channel information of these transmissions. Hence the basestation can perform some degree of interference cancellation and even perform Maximum Ratio Combining. Similarly, in the downlink the basestation can coordinate transmissions to achieve transmit diversity gains and avoid co-sector interference. Hence we can ignore intrasite interference issues and focus on inter-site interference. Let us first consider the downlink case. We assume that each basestation uses the same transmission power for each downlink resource such that if all downlink resources are utilized then the full downlink power is also used. If each basestation knows which resources are being utilized by its neighbours (through information sharing) then it can easily compute the interference experienced by each user given the user’s channel gains to neighbouring sectors. This channel
information is obtained through pilot measurements used by handover algorithms. This interference information can then be used to determine how many resources should be used for the transmission to maintain the flow’s QoS constraints. However in the case of the uplink this interference computation cannot be so easily performed because the source of the interference (the users that transmit in neighbouring sectors) changes with each frame and also the transmission powers of those users may also change and so we take a different approach. Since future traffic will be primarily video in nature, assume that each flow needs to maintain a certain rate with given PDB and PLR constraints. For a given uplink channel gain g, uplink bandwidth b and transmission power p we have a Shannon rate of pg r = b log2 1 + I + σ2 where I is the interference and σ 2 the noise, experienced over the transmission bandwidth. Since only users near the cell edge cause significant interference then for such transmissions (i.e., those with sufficiently small spectral efficiencies) we can make the assumption that bpg r= . I + σ2 The interference density caused by this transmission to a neighbour is given by pˆ g where gˆ is the channel gain of the user to this neighbour. Therefore the total interference caused to the neighbouring sector is given by pˆ gb =
rˆ g (I + σ 2 ) g
which is independent of the transmission power and bandwidth used. Hence the total interference caused to all neighbours is constant no matter what power control algorithm is used. Hence in this case interference management cannot be used to reduce interference and a more appropriate objective would be to shape the interference caused so as to minimize its variation and hence reduce the channel estimation error. Given this new objective one can design a suitable power control algorithm. However we do not provide details due to space limitations. The interference experienced by a sector increases as the loads in neighbouring sectors increase. Present simulation results are typically based on the assumption of full buffer traffic with no QoS constraints. The objective is then to maximize sector throughput while achieving acceptable performance for cell edge users. However, under this assumption, interference issues tend to be exaggerated. The reason being that, with QoS constraints, the system cannot run at full utilization and hence the average interference caused to neighbours is reduced. This effect is illustrated in Figure 7. Hence inter-site interference issues will tend to be of less importance when QoS is taken into account and simple approaches can be used to reduce each neighbour’s loading in the rare cases where it becomes an issue. However, we should also note that if, as previously suggested, excess capacity is to be used to provide service to flows that have exceeded their
Sector Capacity
Full Buffer
Flow Queues
Non−QoS Service PDB & PLR
Frame Usage
QoS Service Fig. 7.
Channel Utilization for Full Buffer and QoS Traffic
GBR then an increase in interference is expected. Hence the degree by which this use of excess resources is followed will determine the potential for interference issues. Conclusion: For QoS services, the power allocation objective should be the reduction of interference variation (as opposed to interference reduction). Full buffer, delay tolerant traffic tends to exaggerate the importance of intersector interference C. Load Management Load management entails various algorithms such as congestion control, admission control and load balancing. With best effort (full buffer) traffic assumptions, the objective is simply to maximize sector throughput while achieving some arbitrary fairness (e.g., Proportional Fairness). However with QoS services the objective now becomes maximization of the total number of users in the system while maintaining the QoS constraints of all flows of all users. Hence load management is of greater importance since there is no longer the ability of simply adjusting one’s throughput to accommodate additional users in the system or changes in sector capacity caused by variations in channel conditions. In order to provide load management one must first be able to accurately measure load. With voice service, the load incurred by each user is typically well defined and near constant (e.g., 12 kbps when in the active state) and so the number of active users is a good measure of sector loading. However in the case of data there is a large variation in the resource needs for each flow and hence the number of flows does not necessarily indicate the true loading. Another approach would be to define load based on resource (power or bandwidth) utilization. However, this metric does not take into account the priority of the active services. For example, if there is a single best effort flow then it can potentially use all radio resources. However this does not constitute heavy loading since the flow can be reduced to allow other flows into the system. Hence a loading metric must also take into account the QoS of the flows. With a utility-based scheduling framework, a gradient ascent algorithm is used to make allocations. In fact what is termed the priority of a flow is actually the gradient of the utility of
the flow with respect to allocated resources. In steady state all of these gradients approach a fixed value. As the system becomes more loaded, flow delays start to increase and the corresponding gradients also increase. Therefore the steady state value of this common gradient (i.e., priority) provides an indication of loading. We define the load of the system as the, suitably filtered, priority of the last flow to be scheduled in the frame. If a flow’s queue becomes empty then the gradient is zero (i.e., the utility does not increase as additional resources are assigned). The priority approaches infinity as the loading approaches the system capacity since the utility gradient goes to infinity. This is because the associated performance metric (e.g. delay) approaches its threshold value at high loading. Given this definition of loading we can now specify appropriate algorithms for congestion and admission control. Due to space limitations we do not provide details. Once the load exceeds a first threshold then admission control is invoked and an incoming flow request is blocked with some probability based on the class of the user. When the load exceeds a second (higher) threshold then congestion control is invoked and appropriate actions are taken (e.g., reduction of GBR values for lower class users, termination of flows etc.). Conclusion: The importance of load management increases as QoS services are introduced. Such management requires an appropriate load metric (as proposed here) since those based on resource utilization or number of active users do not properly reflect loading VI. S UMMARY AND F UTURE W ORK In this paper we considered the trending of the traffic distribution for wireless data and illustrated that pricing together with optimal radio resource management will become of even greater importance in the near future. We then provided an overview of the framework that will be used for QoS services. Given this framework and the fact that future traffic will be provided with QoS guarantees, we suggested a simple approach to pricing of such services. We then showed the impacts on radio resource management algorithms and provided a brief summary of how these may be handled. More detailed studies of these proposed solutions will be performed in the near future. Our intent was to show that present assumptions may lead to poor algorithms for future services. R EFERENCES [1] H, Holma and A. Toskala, “LTE for UMTS - OFDMA and SC-FDMA Based Radio Access”, Wiley (2009). [2] J. Andrews, A. Ghosh and R. Muhamed, “Fundamentals of WiMAX: Understanding Broadband Wireless Networking”, Prentice Hall (2007) [3] “Cisco Visual Networking Index: Global Mobile Data Traffic Forecast Update, 2009-2014”, Cisco Whitepaper, Feb 2010, www.cisco.com. [4] “Mobile Broadband Capacity Constraints and the need for Optimization”, Rysavy Research, February 2010. [5] “ Technical Specification Group Services and System Aspects; Policy and Charging Control Architecture (Release 8)”, 3GPP TS 23.203 v8.9.0, March 2010. [6] P. Hosein, “QoS Control for WCDMA High Speed Packet Data”, IEEE Conference on Mobile and Wireless Communications Networks, Stockholm, Sweden, Sept. 2002. [7] P. Hosein, “Capacity of Packetized Voice Services over Time-Shared Wireless Packet Data Channels”, INFOCOM, Miami, USA, March, 2005.