Alcatel•Lucent

Analysis of the impact of traffic growth on the evolution of access

Corporate CTO Organization, Bell Labs, Alcatel-Lucent

Executive Summary The growth of IP traffic per subscriber on fixed and mobile networks is growing at rates of 30-100% per annum due to the development and use of: i) manifold new web applications; ii) the increase in the number and variety of IP-enabled devices; and iii) the consumed per application. However, the average revenue to the provider per subscriber of their broadband access service is approximately flat for both fixed and mobile networks. Thus it would appear that there is an essential economic conundrum in which the investment in infrastructure required to support the continued innovation and expansion in access to such web services by more users on more devices may not be sustainable. In this paper, we investigate the validity of this argument by comparing the cost of the network expansion required to match the traffic growth with the revenue potential projected by simple extrapolation from current trends. We find that indeed this dichotomy is real, with the cost of network expansion per subscriber exceeding the revenue per subscriber in the next 2-3 years, for both fixed and mobile networks. There are two clear solutions to this conundrum: a) to invest at a level less than required to support the traffic growth that would otherwise occur; or b) to increase the revenue associated with the anticipated traffic growth by an increase in broadband services adoption. We argue that the latter is clearly preferable from the standpoint of the subscriber, the network operator, and the application and content providers, as it maximizes the subscriber experience and the innovation potential associated with the Internet and associated IP services. We also propose a mathematical model that shows that the best effort services actually benefit from the existence of the managed services, so that the two can be regarded as ‘symbiotic’ in nature.

Copyright © 2010 Alcatel-Lucent Page 1 Introduction While the growth in consumer traffic has been remarkable in recent years − one mobile operator saw 5000% Compound Annual Growth Rate (CAGR) in three years − the reality is that the potential for higher growth is still to come. Consumer traffic has been driven in large part by the demand for video-related services, such as web video streaming and video file sharing or downloading. By all accounts, video represents the lion’s share of the traffic on the Internet today if all forms of video are considered (including video files downloaded via P2P or HTTP) as shown in Figure 1[1]. The bandwidth generated by increases each year as broadband access speeds have increased and the average quality of the subscriber experience consequently improves. In addition, many content providers are providing options that move beyond streaming to the PC (with the typical low-to-medium quality experience associated with those services) to streaming directly to the TV through Internet-video enabled Set Top Boxes (STBs) and soon to broadband enabled TVs, whose numbers are predicted to rise from 130 million in 2009 to more than 360 million in 2014 [2].

(peak) traffic breakdown: downstream

HTTP (video) 15.3% Flash / RTSP video 49.4% P2P 18.0% other

14.8%

2.5%

Figure 1. Analysis of the relative contribution of video to Internet traffic [1]

Figure 2 illustrates the disparate bandwidth requirements of a variety of different video services that can and will be delivered over the Internet. Given that the quality of Internet video today is well below that of broadcast video, the potential remains high for even greater traffic growth. For example, a single one hour broadcast quality High Definition (HD) video application consumes as much network bandwidth as 150 ten minute web-quality videos or 600 ten minute Voice-over-IP (VoIP) phone calls. Today cable, satellite, and providers are offering HD at 720p (which equals 720 lines of horizontal resolution, with 60 full frames per second). However, the demand for full 1080p resolution is growing to match the video quality of Blue-Ray DVDs and the 1080p content offered for download by online content providers. As shown in Figure 2, this alone will double the bandwidth of today’s broadcast quality HD. Beyond 1080p there is now strong momentum around 3D HD technology, with ESPN and other broadcasters announcing that they will offer 85 different sporting events in 3D in early 2010 [3]. 3D HD will act as a further bandwidth multiplier (with the precise multiplier depending on the methodology and the number of viewing angles embedded) further increasing network demands by as much as an additional 60- 100%. It appears that there is no end to the consumer appetite for a better, more immersive content experience – an experience that is ‘as good as, or better, than being there’.

Copyright © 2010 Alcatel-Lucent Page 2 One potentially offsetting factor is the improvement in video compression technology that typically occurs over time, with, for example, MPEG2 1080i HD video requiring approximately 20Mbps, whereas the newer H.264/VC-1 (MPEG4 part 10) standard requires only approximately 6-9Mbps to encode the same content without visibly inferior quality. This effect is represented in Figure 2, in the form of the bandwidth demands projected for 2014. It is apparent that, at best, the bandwidth requirement per stream will be halved, but with the number and resolution of video-related applications consumed by subscribers expected to increase significantly with the continued expansion of ultrabroadband networks with bandwidths in excess of 25Mbps per subscriber (which today comprise only approximately 13-16% of broadband connections [4]), the gain due to increased compression efficiency will not significantly alter the essential bandwidth growth trajectory.

Figure 2. Comparison of bandwidth demands of different Internet services

Last, the growth in the number of IP or Internet-enabled devices must be considered, since this serves as an additional bandwidth multiplier that represents the growth in subscriber demand, as subscribers evolve from their current Internet-usage patterns to more advanced usage. In Figure 3, we show aggregate data for the anticipated growth in fixed and mobile IP devices in the United States over the next few years.

Copyright © 2010 Alcatel-Lucent Page 3 Figure 3. Growth in the number of Internet-enabled devices in the US. (The non-Handset Mobile device category includes laptops, netbooks, tablets, e-books and portable gaming consoles. The Fixed IP device category includes laptops, VoIP phones and IP STBs). Source: [5]

Referring to Figure 3, it is apparent that 100s of millions of new fixed and mobile IP-enabled devices will become active in the US in the next few years, representing approximately 2 new devices for every resident. This is therefore a greater compounding factor to the bandwidth demand than the drive towards higher resolution video, which will likely benefit from the improvements in compression technology as outlined above. In this paper, we build subscriber traffic growth models for fixed and mobile networks using a “bottoms up” approach that takes into account subscriber types, the applications they use, and the required “busy hour” dimensioning. These traffic forecasts are then used to dimension the respective fixed and mobile networks to evaluate the net impact on the network capacity. This, in turn, is used to quantify the cost of the required network expansion, for both fixed and mobile networks due to the traffic growth. Finally, we evaluate the consequences of this significant growth in services on the end user experience, if all services are treated equivalently and if no network capacity expansion occurs, and contrast that with the manifest benefits of the converse: services differentiation and attendant network capacity expansion.

1. Fixed Network Traffic Growth Determining a representative network traffic profile is challenging as there are potentially many types of users using many different applications. To this end, our analysis of consumer traffic growth will start with the assumption in a recent Instat study [6] that there are three types of user or household profiles: Passive, Social, and Power or Early Adopters (or A, B, and C). We will also adopt the partitioning of consumer entertainment usage behaviors in that study as our starting point: Passive users represent 47% of consumers, 41% are Social users, and 12% are Power users.

Copyright © 2010 Alcatel-Lucent Page 4 Next we will consider the various traffic sources (applications and content) that will be used, the bandwidth associated with each, and the multiplex of applications associated with each user type. Since we are concerned with the traffic per home, we must account for multiple concurrent application consumption due to multiple users per home. Different analysts have communicated a wide range of consumer traffic models, in part due to unique demographics.Regional differences can also play a part, e.g., significantly higher percentage of European households watch online video than US households, probably due to the high penetration of cable/satellite (high quality video) in the US (above 80%) compared to Europe, which has average cable penetration rates below 20% [7]. Table 1 illustrates bandwidth requirements of different Internet applications, and the association with each user type: A, B, and C. Based on the applications selected, the peak bandwidth requirements for each user type are calculated. This represents the maximum bandwidth assuming that all applications are accessed concurrently. This essentially assumes that households have multiple Internet users, or that some applications run as background processes (e.g., Peer to Peer (P2P) or scheduled updates, backups or content downloads) resulting in a high level of application coincidence or concurrency. In reality this does not occur, and so the distribution of the application usage over time needs to be considered, which we consider in the following section, using ‘busy hour’ dimensioning.

Table 1. Consumer Traffic sources and example of peak bandwidth requirements

Type A Type B Type C Downlink BW, Application Desired Quality Uplink BW, Kbps Household Household Household Kbps (Passive) (Social) (Power) VoIP 3.92 MOS 32 32 ü ü Video Calling Standard Def 1,000 1,000 ü ü Audio Streaming Standard Def 67 0 ü ü Web Radio Standard Def 67 0 ü ü Web Video Streaming Small window 403 0 ü ü Web TV SD Standard Def 1,000 0 ü Broadcast quality TV HD High Def 6,000 0 ü Interactive Gaming Premium 3,000 85 ü ü Instant Messaging No options 1 1 ü ü ü Web Browsing Standard 1,000 1 ü ü ü Peer2Peer No options 500 500 ü ü Email No options 0.040 0.040 ü ü ü Worst Case Peak Bandwidth Requirements 1,472 7,070 11,600

Even from this simple analysis, it is apparent that there is a potentially an almost eightfold difference in bandwidth consumption between type A users and type C users (1.5 Mbps compared to 11.6 Mbps). This difference is due in large part to the differences in the amount and type of video consumed. Table 2 illustrates the expected migration of users towards an increasingly heavy consumption pattern, representing the evolution in number of devices (Figure 3) and application usage projected above. The assumption we make to derive the future profile mix is that a proportion of passive users become social users over time and the social users of today become power users in the future. This is a reasonable assumption if we look at the recent past: for example, Internet video on the TV was non-existent two years ago but according to InStat Online Entertainment [6], 30% of broadband homes have now viewed online video on the TV. Another Internet video service, Netflix, has 11 million subscribers in the US (less than a decade after launch) and 40 % of their customers watch online streaming video to the TV or PC, consuming a minimum of 1.5 Mbps, with 4 Mbps or more desired for an optimum experience.

Copyright © 2010 Alcatel-Lucent Page 5 These user/household profiles will be used together with the bandwidth requirements per profile, and the busy hour dimensioning to determine the actual aggregate bandwidth requirements in the network.

Table 2. Evolution of user type distribution

User Definitions Distribution 2009 Distribution 2014 Passive User 47% 30% Social User 41% 50% Power User 12% 20%

1.1 Network Dimensioning Based on Busy Hour Network operators do not dimension their networks by calculating the simple sum of all bandwidth required across all users and user profiles (the analysis presented above) assuming 100% concurrency of usage, but rather they evaluate and dimension the network based on the ‘busy hour’ usage – i.e., the bandwidth required at that point in the day when the traffic is heaviest, which typically occurs in the ‘prime time’ window between 6pm and 10pm. To turn the above consumer traffic profiles into busy hour traffic on which we can dimension a network we therefore combine 3 elements: user profiles, network application bandwidths, and busy hour dimensioning. To calculate the contribution of users’ traffic to the busy hour, the bandwidth per service is divided by the period over which the service is used. For example, if we assume an average user watched 20 minutes of online video a day at 400 kbps (PC Quality video), and that occurs during a four hour window in the evening, the contribution to the busy hour is: 20 mins/4 hours multiplied by 400 kbps or 33 kbps. In essence, this is the statistical gain due to different users logging on and consuming services at different times of the day. However, such statistical gains are likely to diminish as the consumption of long form (20+ minute TV-type to multi-hour movies) video increases. To calculate busy hour dimensioning we need to know the total number of minutes for each application and the bandwidth associated with each application. Table 3 is our 2014 projection of daily minutes per Household (HH) for key Internet services for a moderate growth projection.

Copyright © 2010 Alcatel-Lucent Page 6 Table 3. Forecast minutes and bandwidth consumed in 2014 (Moderate Growth Profile)

Moderate Min per HH Kbps per Application/BH Services Kbps Passive Social Power Passive Social Power VoIP 32 0 10 15 0 1 2 Video Calling 1000 0 20 30 0 83 125 Audio Streaming 67 30 30 0 8 8 0 Web Radio 67 0 30 60 0 8 17 Web Video 403 15 30 30 25 50 50 Web TV SD 1500 5 20 30 31 125 188 Web TV-HD (10% 3D) 9900 0 0 105 0 0 4331 Interactive Gaming 3000 0 15 30 0 188 375 Instant Messaging 1 5 5 5 0 0 0 Web Browsing 1000 12 18 24 50 75 100 Peer2Peer/HTTP Download 500 9 30 45 19 63 94 Email 0 4 6 8 0 0 0 Average Bandwidth per Household 134 602 5282 Weighted Average Bandwidth/Household 1397 Video Minutes per Household 29 80 210 Weighted Average Video Minutes/User 45 For this calculation, we assume multiple concurrent users per home (2) and the distribution of user types described in Table 2. The bandwidth per application is multiplied by the total minutes of usage per application and per user type. We then we assume a four hour evening window in which users consume the majority of their fixed network applications/content, so that the total minute usage is divided by 4 hours (or 240 minutes) to determine the contribution of that application to the busy hour. Then using the distribution of different user types we can calculate the weighted average bandwidth (1400 Kbps) and video application minutes. Using a similar methodology we have calculated an aggressive bandwidth growth that results in a value of 4300 kbps per household and 79 minutes of video application usage per user. Figure 4 summarizes our forecast for subscriber busy hour traffic, for the two different traffic growth models. As indicated above, the moderate growth rate traffic profile reaches 1400 kbps per household, and the high growth rate profile reaches 4.2 Mbps per household in 2014. Using these two data points one can calculate the compound annual growth rate of traffic from the values observed currently (in late 2009). In this way we find a CAGR of 67% for the moderate projection and 100% CAGR for the high growth rate. A September 2009 comScore [8] survey found that the average user watched 9.8 hours of on line video that month, representing a 100% CAGR from the previous year. We note, therefore, that our aggressive growth rate corresponds to the growth rate reported by comScore for the past year, so it may in fact be a conservative estimate in reality, given the manifold increase in services and bandwidth per service anticipated in the introduction.

Copyright © 2010 Alcatel-Lucent Page 7 Figure 4. Traffic forecast in Fixed networks (Moderate and Aggressive Growth models).

1.2 Mobile Network Traffic Growth With the emergence of easy-to-use “,” the mobile data subscriber’s usage has grown from minimal to levels that could ultimately approach that of fixed networks, as applications that were initially targeted at fixed users (e.g., web browsing, video streaming, gaming, video calling) are now commonplace on the mobile terminal. Table 4 illustrates the different applications and associated bandwidths for mobile networks and the three different user types as previously defined. Compared to the fixed network, the required bandwidths are 7-14 times smaller on a mobile network, primarily due to the smaller mobile device screen sizes, which reduce the bandwidth required per video stream for mobile terminals. However as mobile users begin to use 3G and LTE modem cards on their laptops, there is the potential for fixed network levels of video traffic on mobile networks in future; this additional growth potential is not currently captured in our model.

Table 4. Mobile Consumer Traffic Sources and Bandwidths

Application Desired Quality of Downlink BW Type A User Type C User Type B Social (non-Voice) BW (Kbps) Passive Power

VoIP Good 12 Push2Talk Excellent 15 ü ü Audio Streaming Good 16 ü ü Mobile Web Radio Excellent 128 ü ü Video Streaming Good 192 ü ü HD Video Streaming Excellent 578 ü Video share Excellent 128 ü Interactive Gaming Low BW 25 ü ü Instant Messaging No options 1 ü ü ü PC Data Card Medium 800 Web Browsin/Email Medium 240 ü ü ü Peak Bandwidth Requirements 241 617 1,323

Copyright © 2010 Alcatel-Lucent Page 8 To create a traffic model we again assume three broad categories of users (Passive, Social, and Power), just as in the fixed network case. Also, as in the fixed network case, we assume that today’s Passive users will become Social users over time and today’s Social user will become tomorrow’s Power user, with slightly higher levels of Power users (30%) compared to the fixed case (20%) and a concomitant decrease in the fraction of Passive users (20%) in the mobile case, relative to the fixed case (30%). To compute the busy hour bandwidth for mobile subscribers, we will take an approach similar to that for the fixed network case. Notably, for mobile networks, it has been observed that the traffic profile is approximately flat from about 9am to 9pm (a total of 12 hours), meaning that mobile users use their device continuously throughout the day and there is no ‘prime time’ period as for the fixed network case. Table 5 summarizes the results for both the moderate and high growth rate cases of mobile users’ traffic consumption in 2014.

Table 5. 2014 Mobile Traffic Consumption per Subscriber (Moderate and Aggressive growth)

Moderate Traffic Profile Passive Social Power BH BW (Kbps) 1.0 9.3 18.8 Sub Airtime Usage (mins) 14 55 101 Usage Mix 20% 50% 30% Weighted Average Peak BH BW (Kbps) 10 Total Usage per Day (minutes) 43 Content consumed (MB/Mo/Sub) 1,633

Aggressive Traffic Profile Passive Social Power BH BW (Kbps) 2.8 25.9 52.5 Sub Airtime Usage (mins) 28 109 202 Usage Mix 20% 50% 30% Weighted Average Peak BH BW (Kbps) 29 Total Usage per Day (minutes) 121 Content consumed (MB/Mo/Sub) 7,016

We can see that the busy hour traffic per subscriber is smaller than the fixed network due to the extended period over which mobile users access the wireless network and the lower bandwidth per application due to lower screen resolution. We have also calculated the consumption of capacity in way that is commonly used in reporting mobile data service usage, using MB/month (per subscriber). Industry analysts expect tremendously aggressive traffic growths as smartphones proliferate and new applications emerge, with traffic usage doubling multiple times a year. In December 2009 Yankee Group forecast mobile data traffic growth of 29 times (2900%) by 2015 [9]. Morgan Stanley reported that ATT traffic grew 50 times in the past 3 years and projects a 129% CAGR [10]. Once again, the near universal assumption is that today’s casual user will become tomorrow’s heavy user. This trend is already beginning to emerge with the increasing adoption of smartphones; Roger Entner, head of telecom research for Nielsen, reported that the typical customer consumes about 40 to 80 MBs of wireless capacity a month, whereas the typical iPhone customer uses 400 MB a month (nearly ½ GB)[11]. Using the methodology described, we again create two different mobile user traffic models; moderate and aggressive growth rates as shown in Figure 5. The moderate growth rate curve reflects a weighted average busy hour contribution of 10 kbps per subscriber and a total of 43 minutes/day airtime in 2014. This results in a CAGR of 62% which, although seemingly high, is significantly lower than that forecasted by Morgan Stanley, Verizon, and others. The aggressive growth rate model assumes a busy hour contribution of 30 kbps at 121 minutes/day of airtime in 2014 – which results in a CAGR of 114%. We note that this value is still smaller than the Morgan

Copyright © 2010 Alcatel-Lucent Page 9 Stanley forecast of 129%, so our models may in fact be conservative. As would be expected, very similar growth rates are computed when using the metric of MB/mo/subscriber (also shown in Figure 5).

Figure 5. Traffic Forecast for Mobile Subscribers; Left – Busy Hour, Right – MB/Sub/Mo.

2. Network Impact from Consumer Traffic Growth

2.1 Impact on Wireline from Traffic Growth Fixed network topologies are typically a tree structure with several traffic aggregation points between the subscriber and the hub office (the point at which traffic typically leaves the operator network and traverses the wider Internet). Figure 6 shows a typical fixed operator IP network with Digital Subscriber Line (DSL) Access equipment (labeled ‘D’), either in cabinets closer to homes or centrally deployed in the end office. The traffic from the access nodes is aggregated through Layer 2 nodes (‘L2’) and then Layer 3 nodes (‘L3’) with multiple aggregation levels depending on the size of the metropolitan area served. It is important to note that although the end user may have a connection that supports 20Mbps or more in the ‘first mile’ (from the access , D, to the home), the multiple levels of aggregation result in only approximately 100 kbps per user on average of Internet traffic at the hub office. As indicated earlier, this busy hour dimensioning is possible due to the statistical of service usage.

Copyright © 2010 Alcatel-Lucent Page 10 Figure 6 Schematic depiction of a typical fixed operator network

2.1.1 Impact of Traffic Growth on Fixed Network To compute the net cost (operational and capital) to the network operator of the predicted traffic growth, we assumed a Fiber to the Node (FTTN) access architecture and used typical cost numbers for FTTN network [12], and used a 10 year cost depreciation model to map these costs to the period in question (2009-2014). A second consideration for network operator is the cost of connecting their network to the wider Internet, typically termed transit costs. In the US the NANOG (North American Network Operators Group) reports a transit cost of about $25/Mbps/Month, so that a gigabit connection to the Internet costs $25000/month. We assume that this decreases at 10% per annum over the 2009 to 2014 period in question. Pyramid Research [13] has forecast the Average Revenue per User (ARPU) for High Speed Internet users in North America will remain flat to slightly decreasing at around $30/month, as shown in Figure 7, along with the calculated per-subscriber costs due to traffic growth for the moderate and aggressive growth scenarios. It is immediately apparent that the average network cost per subscriber associated with the moderate growth curve approaches the revenue per subscriber by 2013, and the high growth scenario exceeds the average revenue per subscriber by the end of 2011, which is clearly a difficult economic scenario to sustain.

Copyright © 2010 Alcatel-Lucent Page 11 Figure 7. Projected fixed network revenue and costs per subscriber

2.2 Impact of Traffic Growth on Mobile Network Similar to the fixed network analysis, the mobile network model used leverages typical industry average data for network costs. The approach is a based on a holistic analysis of leading US mobile operators, using publicly reported financial data and other publicly released network, subscriber, and traffic data, and reflects a mix of 2G and 3G networks [12]. We estimate the revenues from subscriber data services from published data offers and the subscriber statistics. The cost of the network capacity takes into account all network operation, interconnect, roaming, facility leases, tower leases, power, and capital costs. We allocate costs due to data services in proportion to the total {data + voice} capacity required.

Copyright © 2010 Alcatel-Lucent Page 12 Figure 8 Schematic depiction of a typical 3G/LTE mobile operator network

The analysis assumes a 15% reduction in Cost/MB per year for the next 3 years and 20% per year after that. The reduction is due in part to more subscribers and the advent of LTE, which will start deployment in earnest in the next 2-3 years. LTE brings a 3-4 times bandwidth improvement and thus helps lower the cost per MB for the access portion of the network. Figure 9 provides a summary of the projected mobile network scenarios, showing the monthly ARPU per data subscriber, the monthly cost per data subscriber, and the two traffic growth curves. Using the moderate traffic growth assumption (46% CAGR) the costs exceed revenue by the end of 2012 (less than 3 years away); with the aggressive growth profile this point occurs by the end of 2010. These trends are similar to those described for the fixed network, although the problem is clearly even more severe for mobile networks. We also note that the rate of divergence of the cost and revenue curves for the aggressive growth case is manifold higher in the mobile network case, in line other industry analysts projections [9-11]. In summary, we conclude that the cost of building out fixed and mobile networks will be prohibitive, if the revenue model remains flat. The choice the network operator faces are: i) to limit the capacity expansion to something less than that required by the actual unconstrained bandwidth demand; or ii) to increase the revenue that is associated with, and will drive, the capacity expansion. In the next section, we will highlight how the former approach (limiting the capacity expansion) will result in an unacceptable level of quality of service and the clear benefit to all if an additional premium quality of service tier or capacity is offered.

Copyright © 2010 Alcatel-Lucent Page 13 Figure 9. Projected Mobile network revenue and costs per subscriber

3. Impact of Network Congestion on Innovation In addition to slowed web page downloads, freezing video streams, dropped VoIP calls caused by limited network capacity, there is yet another way in which constrained capacity will impact applications: increased delay or jitter. Jitter is simply the variation in the delay, which is the average time taken for data packets to appear, following a request. Real time applications such as video conferencing, voice services, streaming video, Internet gaming and cloud computing are all sensitive to delay and jitter, as they require ‘round trip times’ of 100ms or less for the delay not to be perceptible to the end user [14]. In particular, innovation in cloud computing is really in its infancy and will impact every segment of residential and business applications as standard desktop and enterprise applications move into the cloud data centers in order to minimize the operational (IT) and capital (server upgrade) costs in the premises and create dynamic new business models. But, in order for this to be realizable, the delay in communication between the end user terminal and the application running in the cloud must be imperceptible from that experienced by the user when the application is locally hosted. This results in a strong requirement for minimized network delay, as well as variation in that delay (jitter). We now investigate the impact of network congestion on delay and jitter. We will apply the traffic model of Norros [15] to illustrate the relationship between the aggregation of so-called ‘bursty’ traffic sources and delay/jitter. The basic model is shown in Figure 10. We assume a line rate of 10Mbps and 20 bursty variable bitrate traffic sources with a mean of 400 kbps each, aggregated within this pipe. We increase the number of traffic sources (N) from 1 to 20 and look at the impact on delay and jitter. Secondly, we look at the impact of moving some of the traffic sources to a new enhanced QoS pipe.

Copyright © 2010 Alcatel-Lucent Page 14 Figure 10. Traffic Model Assumptions

From the left panel of 11, it is clear that as the number of traffic sources (N) increases from 1 to 20, the jitter rises steeply at around N = 15 (Average BW/Line rate = 0.6). In other words, as the bursty traffic becomes greater than 60% of the line rate, the jitter grows by orders of magnitude, rendering the network essentially useless for delay-sensitive services. Using the same traffic model as in Figure 10, we now add additional capacity that is associated with a premium quality of service. We further assume that only 5 users (of the 20) take advantage of this service. The right panel of Figure 11 shows the impact on the jitter. It is immediately apparent that when the 5 users are moved to an advanced service pipe there is considerable improvement the jitter that the remaining 15 users (those that did not participate in this service) experience, as the jitter experienced drops by a factor of 30, from 4s to 124ms, which is acceptable for many real time services. Lastly, we can see that if an additional 4.5 Mbps is given to premium service tier, the 5 premium users will have jitter well below 100ms under all circumstances. This service guarantee could be offered for a fee (determined by the subscriber’s willingness to pay), as is the case for all such ‘managed services’ in today’s network (e.g. Voice and IPTV services).

Figure 11. Evaluating the impact of Network Congestion on delay

Copyright © 2010 Alcatel-Lucent Page 15 The conclusion is quite clear even from this simple example: that there is a net benefit to all parties in offering a premium or managed service. The benefit is actually threefold:

i) The reduction in network congestion associated with the creation of a managed service translates to reduction in the jitter to the point where advanced real time services can be offered to all users with an acceptable average service experience.

ii) The managed service itself further benefits subscribers looking for additional performance guarantees (beyond an acceptable ‘average’ experience).

iii) The managed service will also enhance broadband services adoption, thereby supporting the required capacity expansion, which will benefit all users.

To further investigate this connection between revenue-generating managed services and the concomitant improvement in best effort services, we have developed [16] a model of a dynamical economics system in which the main players are subscribers, network providers and application/content providers, with the key goal to track the evolution of the consequences of their decision-making over time. We have created two independent models, the first in which there is only ‘Best Effort’ Internet service and a second in which Best Effort and Managed Services co- exist; the intrinsic behavior of the Best Effort subscribers is common to both models.

An essential element in these models is that some fraction of total network revenue from each time period is applied to capital expenditures for enhancing network infrastructure in the subsequent period. Under a broad set of parameters, the natural consequence is growth over time of network capacity. In both models, Best Effort subscription prices are held fixed and the number of subscribers to the service is the equilibrium implied by the monotonicity of the willingness to pay for Quality of Experience, which in turn depends monotonically on per-subscriber network capacity. For Managed Services the demand is obtained from a constant elasticity model in which the price is obtained by an optimization performed jointly by the network provider and the application/content providers. In both scenarios, the initial network size is the same and the operator can invest up to 25% of its current revenue to add and maintain network capacity for the next period. We also assume that the cost to add and maintain network capacity is dynamic and decreases by 5% in each period.

In short, what we find is that since capacity expansion is constrained by a network operator’s resources (which ultimately depend on the revenue stream), offering Managed Services generates additional revenue, which gives the operator a larger budget for network expansion. Most notably, the users of the Best Effort service also benefit from this expansion, as is demonstrated by a comparison of the two scenarios and shown in Figure 12 below. The left panel clearly shows the growth in overall network capacity and the right panel shows the concomitant growth in the Best Effort capacity that is supported by this new revenue and investment.

Copyright © 2010 Alcatel-Lucent Page 16 Figure 12. Increase in network capacity over time for two scenarios: Best Effort services only and Managed + Best Effort Services. Left: Total Network Capacity; Right: Best-Effort Capacity

4. Conclusions In this paper, we evaluated the cost of the capacity expansion required to meet the anticipated bandwidth demands for moderate and aggressive growth rate cases and compared these costs to the attendant revenue, when viewed as a simple extrapolation from the current revenue. We find that there is a dichotomy, with the cost of network expansion per subscriber exceeding the revenue per subscriber in the next 2-3 years, for both fixed and mobile networks – a situation that is clearly undesirable from an economic standpoint. There are two clear solutions to this conundrum – to invest at a level less than required to support the traffic growth that would otherwise occur, or to derive more revenue from the ensuing traffic growth. We argue that the latter is clearly preferable from the standpoint of the subscriber, the network operator and the application and content providers (ACPs), as it maximizes the subscriber experience and the innovation potential associated with the Internet and associated IP services. Furthermore, we show that a model whereby a combination of managed services and best effort services co-exist and there is enhanced revenue potential associated with the former, results in a highly desirable outcome in which the network capacity expands, and to the benefit of both service types.

5. References

[1] Alcatel-Lucent measurements of high speed Internet traffic running over operator networks, June to December 2009. [2] DIGITAL TV BUSINESS, TECHNOLOGY & MARKET RESEARCH NEWS, “Broadband-Enabled TV Households to Top 360 Million by 2014”, January 15th 2010, http://www.digitaltvnews.net/content/?p=12137 . [3] ESPN.com “ESPN 3D to show soccer, football, More”, http://sports.espn.go.com/espn/news/story?id=4796555 . [4] Alcatel-Lucent analysis of Broadband Trends, Infonetics and Dell’Oro Q2 and Q3 2009 reports.

Copyright © 2010 Alcatel-Lucent Page 17 [5] Alcatel-Lucent analysis of data supplied by Pyramid Research, Ovum and Strategy Analytics. [6] InStat “Competing Business Models for the Future of Digital Entertainment,” December 2009. [7] Solon Management Consulting GmbH & Co. KG, Solon European Survey 2009, Section 1.1. [8] Source: www.comscore.com. [9] Declan Lonergan, “The Holy Grail of Mobile Broadband Pricing”, Anchor Report Yankee Group December 2009. [10] Morgan Stanley presentation on Internet Trends, Web 2.0 Summit, San Francisco, October 2009. [11] Reported by PCWorld in December 2009, www.pcworld.com/article/184589/atandt_iphone_users_irate_at_idea_of_usagebased_pricing.html [12] Model costs and methodology to be published in Bell Labs Technical Journal, 2010. [13] Pyramid Research, US FIXED COMMUNICATIONS DEMAND - RESIDENTIAL - June, 2009. [14] T. Coppens et al, “Access Network Delay in Networked Games”, NetGames03, Redwood Coty, CA (2003). [15] Ilkka Norros, "A storage model with self-similar input", Queuing Systems, Volume 16, Numbers 3-4, pp. 387-396, September 1994. [16] Bell Labs Mathematical Sciences research, Bell Labs Technical Journal, in preparation 2010.

Copyright © 2010 Alcatel-Lucent Page 18