Watching a Nation Watch IPTV

Paper ID: # 65

Meeyoung Cha∗ Pablo Rodriguez¶ Jon Crowcroft† Sue Moon‡ Xavier Amatriain¶

∗MPI-SWS ¶Telefonica Research †University of Cambridge ‡KAIST

ABSTRACTFor half a century, television has been a dominant and perva-sive mass media, driving many technological advances. Therecent boom of Internet TV (IPTV) will also likely lead toinnovations in the Internet. In contrast to traditional tele-vision, IPTV enables us to monitor the user behavior andnetwork usage of the entire network. The resulting analy-sis provides a clearer picture of how people watch TV andhow the underlying networks and systems can better adaptfor future challenges.

In this paper, we present the first analysis of IPTV work-loads based on traces collected from one of the world’slargest IPTV networks. Our study characterizes severalproperties of viewing sessions, channel popularity dynam-ics, geographical locality, and channel switching behaviors.We discuss the implications of our findings on networksand systems, including the support needed for fast channelchanges. We replicate IPTV on the wider Internet. Our find-ings have important implications on not only existing IPTVsystems, but also the design of the open Internet TV distri-bution systems such as Joost and LiveStation.

1. INTRODUCTIONSince the 1950’s, television has been a dominant andpervasive mass media; it is watched across all age groupsand by almost all countries in the world. Over the years,television has transformed itself into a new media. Thenumber of channels has increased from a few free-to-airbroadcasts to several hundreds for cable, satellite, andInternet TV networks, that transmit more channels toeach user. The video signal itself has changed fromblack & white and color analog to high-quality digitalstream.

Many technological advances were produced by try-ing to meet user needs and expectations in such awidespread media. For example, the large number ofusers that concurrently watch TV sparked the use ofIP multicast in major Internet TV (IPTV) designs tosave on the transmission cost. And now traditionalmedia is converging with newer Internet-based services(e.g., Joost, Zatoo, and LiveStation). In such a con-text, IPTV is a promising starting point for research

because it opens up the door for many innovations whilestill keeping its roots in the traditional TV watchingparadigm.

Despite the widespread usage of television and its im-portance to emerging applications, the ingrained TVviewing habits are not completely understood. NielsenMedia Research [4] spearheads a long-standing researcheffort to estimate TV viewing behaviors through mon-itoring and surveys. However, due to the difficulty ofinstrumenting monitoring devices at individual homes,it is hard to monitor user behavior across the entirenetwork. Sophisticated methods such as stratified sam-pling, systematic sampling, and cluster sampling areused to find a representative set of users and extrapo-late their behaviors to the entire population.

With the recent large-scale deployment of commercialIPTV systems, the IPTV workloads add a new dimen-sion towards understanding TV viewing habits. Thetwo-way connectivity enables us to monitor user behav-ior across the entire network. Therefore, the resultinganalysis provides a clearer picture of how people watchTV. The large pool of users allows us to zoom into dif-ferent groups and combinations of users based on thechannel watched, geographical locations, and viewingpatterns.

In this paper, we present the first in-depth analysisof IPTV workloads using traces obtained from one ofthe world’s largest IPTV networks. Our trace spans asix-month period and a quarter million users, an ordermagnitude larger than that of Neilsen. We characterizeseveral properties of viewing sessions, channel popular-ity dynamics, and content locality, which help contentproducers and service providers create more appealingprograms and a better broadcasting schedule. We alsoidentify the browsing patterns of users to find contentand characterize the user arrival and departure patternsacross channels, which could be used as an input to de-sign future IPTV systems.

The highlights of our work could be summarized asfollows:

1) Over 60% of channel switching events happen within10 seconds, indicating user focus is very volatile.

1

2) Channel popularity changes widely over a day andpopularity of top channels is reinforced duringpeak times. Popularity distribution is Zipf-likefor popular channels and it is truncated for non-popular ones.

3) Whene watching TV, attention span is genre-dependent and ranges from 5 to 21 minutes. Dailyviewing patterns are also heavily dependent ongenre.

4) We observe strong differences in the viewing pat-terns of different regions within a country, wherethe variations are up to 25% across DSLAMs.

5) Channel switching rates depend on the programpopularity, where popular channels have lessprobability of switching. Channel switching isrelatively insensitive to the duration of pro-grams. Channel switching events are predominant(>60%) and can have a major impact in user’sexperience as well as network performance.

6) The set of channels people watch (for longer than 1minute) follows non-sequential channel numbers,while 62% of channel changes are sequential.

Our findings, based on non-sampled data, have sev-eral important implications for the design of both fu-ture television and the open Internet TV distributionsystems. In particular, we find that channel selectionprocess of users can have a major impact on the net-work, as significant channel changes generate unneces-sary network load. Our results also reveals that thedesign of system that supports fast channel switchingand helps users find the right channel quickly is imper-ative to both improving end user experience and mini-mizing the impact in the network. Finally, we find thatthe users’ tastes are not as homogeneous as one wouldexpect over time, geography, and content, and there-fore, our findings can provide important input towardsdesigning an efficient IPTV channel recommendationsystems.

The rest of the paper is organized as follows. Wedescribe the context and background of IPTV in §2.We introduce the dataset and our assumptions in §3.§4 presents our characterization of the viewing session,channel popularity, and geographical locality. §5 fo-cuses on the dynamics of channel switching behaviors.In §6, we discuss the implications of our findings. Wepresent related work in §7, and discuss issues for furtherwork and conclude in §8.

2. CONTEXT AND BACKGROUNDWe introduce the IPTV services and its architecture,and summarize the previous studies on TV habits.

2.1 IPTV MarketThe IPTV market has grown rapidly over the past fewyears. As a whole, an estimated 13.2 million house-holds received IPTV services in 2007 worldwide and thisnumber is projected to grow 52.2% annually through2012 [3]. Major IPTV players are the telcos (telecom-munication service providers) that run national back-bones. Examples are PCCW in Hong Kong, ChinaTelecom in China, KDDI in Japan, France Telecom inFrance, Telefonica in Spain, and AT&T in the UnitedStates. There are other types of IPTV services, thatare not based on high-speed realtime streaming but pri-marily on downloading movies and TV programs beforeviewing (e.g., Apple TV, Vudu, Amazon’s Unbox forTivo). Despite the huge success, no large-scale studieshave been published on the viewing habits from theseservices. In this paper, we study the intrinsic charac-teristics of what and how people watch TV using tracescollected from one of the world’s largest IPTV opera-tors.

2.2 IPTV Service ArchitectureMost telco’s IPTV architecture is based on IP mul-ticast [12, 17]. Figure 1 illustrates a typical servicearchitecture, where customers subscribe to IPTV, IPphone, and Internet access. A network device calledDigital Subscriber Line Access Multiplexer (DSLAM)aggregates traffic from hundreds to thousands of usersand connects to the high-speed IP backbone. ForIPTV distribution, the TV head-end streams all thelive broadcast TV channels towards DSLAMs throughbandwidth-provisioned multicast trees in the backbone.

home

gatewaySTB

PC

TV

DSLAM

customer

premiseTV

head-end

ISP

IP backboneInternet phone

Figure 1: Telco’s IPTV service architecture

Due to limited access bandwidth at the last mile (i.e.,from a DSLAM to the home gateway at the customer’spremise), not all TV channels are delivered to customerssimultaneously. The IPTV network studied in this pa-per streams one channel to each customer. However,multiple channels can be delivered to support multipleTVs at home. The IPTV set-top-box (STB) translatesa channel switching event from a user into a pair of In-ternet Group Management Protocol (IGMP) messages:one to alert the user’s departure from the multicastgroup of a channel and the other to join another mul-ticast group of a new channel. Upon receiving IGMP

2

messages, DSLAMs extend or prune multicast trees dy-namically towards the customers’ STBs.

2.3 Studies on TV ViewingThe Nielsen Media Research [4] provides a wide set ofstatistics on TV viewing and program ratings based ona sample of population. The Nielsen technicians in-stall metering equipment on TV sets at sampled home,and record minute-by-minute viewing of users. Theserecords are collected on a daily basis and mapped withthe broadcast program schedules to produce the NielsenTV ratings. Their sampling devices now gather tracesfrom approximately 25,000 people from 10,000 homesin the United States.

However, there are known issues with sampling meth-ods such as the ones used by Nielsen. Selected house-holds are not accurate in terms of demographic and re-gional weighing. Furthermore, and regardless of theincreasing incentives offered by audience analysis com-panies, over 30% of the selected households refuse to bepart of the studies [18]. There is no available informa-tion of how this may bias the statistics. In contrast,IPTV providers can monitor behaviors of all the usersclosely and continuously without having to install extramonitoring devices.

3. DATASET AND ASSUMPTIONSWe describe our methodology to collect and process logsand discuss our assumptions. Throughout this paper,we interchangeably use the term switching to refer tothe user action of changing channels.

3.1 Trace CollectionWe obtained a collection of IPTV channel switching logsfrom an operational backbone provider. The trace ofover 800GB spans six months from May 1 to October 31,2007, and records the IGMP messages on the channelchanges of 250,000 users. The log includes the following:

• Timestamp in units of seconds

• IP address of the DSLAM

• IP address of the set-top-box (STB)

• IP address of the multicast group (channel)

• Multicast option of join or leave.

Some multicast groups are used to manage set-top boxes(e.g., bootstrapping and software upgrades). We pre-process the logs by excluding events for non-video mul-ticast groups. We then chronologically sort IGMP joinmessages. This allows us to identify which program auser was watching, for how long, and how one switchedchannels. Due to the scale of our trace, we present re-sults that are averaged across the trace period or froma representative day.

There are 150 channels, which are grouped by theIPTV provider as follows: First, there are free-to-airbroadcasted channels within the nation – which we re-fer to as free and local channels produced by 12 re-gional states. Then there are audio, satellite, andpay-per-view (ppv) channels. The rest are cable TV-like content including children’s programs (kids), doc-umentaries (docu), movies (cine), sports, news, andmusic channels. We refer to the collection of comedy,soaps, and reality shows of the remaining channels (e.g.,Fox, Paramount comedy) as mixed.

3.2 Assumptions on User ModesBefore we proceed our analysis and in order to bet-ter understand the user behavior, we will divide useractivity into three different modes, which we will callsurfing, viewing, and away. Our away mode indi-cates whether the user has turned off the TV or the TVis left on without anyone watching it. Our trace doesnot include information about such user modes, becausemost users leave the IP set-top box on and continue re-ceiving multicast stream, even when the television isoff.

To define these three modes, we examine the channelholding times, i.e., the time interval between channelswitchings. Figure 2 shows the cumulative distributionof channel holding times, where the slope of the curvechanges around 10 seconds and again around 1 hour,indicating natural divisions of the times between thesethree different modes. Although these divisions maysomehow look arbitrary, they coincide very well withthose divisions made by the traditional media research1.

1 sec 1 min 1 hour 1 day10 secs0

0.2

0.4

0.6

0.8

1

Channel holding time

CD

F

Surfing Viewing Away

Figure 2: Inferring user modes: surf, view, away

We set the viewing threshold (or viewing period) to 1minute to be consistent with the Nielsen Research andto be able to compare the data better. We set the awaythreshold (or away period) to 1 hour. The most popularchannels in our IPTV system are free-to-air, which arepaid by advertising, and therefore it is hard to find any

1The Nielsen Research demarcates viewing events by theminute [4]. MSK Research in Korea uses 15 seconds forviewing period; TNS Media in Korea considers the channelthat a user spent most time during every 1 minute as theviewed channel.

3

single program longer than 1 hour without having in-terruptions. While some users may watch TV for morethan 1 hour without channel changes, especially whenwatching movie channels, we assume the effect of thoseusers to be negligible as those channels are likely lesspopular. Instead, we focus on the “active online view-ers” – i.e., the users who denote some activity during 1hour period2.

To validate our assumptions, we have also repeatedmost of the antilysis in our paper with different thresh-olds and found the result had very little impact. Due tospace limitations, we do not present those results. How-ever, we believe that the variations on these thresholds,for instance, from 1 minute to 5 minutes or from 1 hourto 2 hours do not change the main results of our paper.

4. HOW, WHEN, AND WHAT PEOPLEWATCH

In this section, we characterize several properties of TVviewing behaviors. In particular, we focus on:

• User session characteristics: when people watchTV and what is their attention span across genre.

• Channel popularity and dynamics: how user inter-ests are spread across channels over time.

• Geographical locality: whether users in the sameregion or DSLAM show similar viewing patterns.

4.1 High-Level Viewing CharacteristicsWe present our findings on viewing characteristics.

4.1.1 Number of Online Users

We examine the number of concurrent online usersbased on our notion of users on- and off-line. Figure 3shows the day-of-week trend over a representative traceperiod. The start of each day is marked in the horizon-tal axis. The number of viewers in Figure 3 shows someinteresting patterns. First, the graph shows diurnal pat-tern with two large peaks around 3PM and 10PM, anda small peak around 8AM. They correspond to dininghours of the population. Other VoD systems showedsimilar diurnal patterns with two peaks [15, 33].

Unlike patterns of Internet usage, Friday and Satur-day consistently had the lowest evening peaks withinthe week. On weekends, the number of viewers rampsup from early morning, which was due to the increasein viewing of children’s programs. Further details re-veal that on weekends the number of distinct viewersincrease by 5% and the total time spent on TV increaseby 30%. Such a trend has been also reported in the sur-vey of the UK workers about their leisure time use [6].2Although we borrow the concept of “active viewer” fromtraditional media research [23], we do not mean to implythat our threshold-based method is a good measure of view-ing activity.

Mon Tue Wed Thu Fri Sat Sun0

20000

40000

60000

80000

Time (week)

Num

ber

of o

nlin

e us

ers

Figure 3: Number of viewers over a week period

4.1.2 Session Characteristics

We describe the users’ session characteristics. An aver-age household watches 2.54 hours and 6.3 distinct chan-nels of TV a day. The average length of time userswatch TV each online session is 1.2 hours, and eachonline session consists of multiple surfing and viewingevents. According to the traditional survey results re-ported by Neilsen [5], an average household watchestelevision for approximately 8 hours and 14 minutes perday and an individual watched 4 hours and 34 minutesa day. Other surveys report that an average householdwatches 3 hours of TV a day [1].

Investigating the session durations more closely, weexamine the times spent on watching television beforechanging a channel (i.e., sojourn times). Using a repre-sentative day trace, we show the frequency of sojourntimes of all online users in Figure 4, in a log-log scale.Significantly greater user activities last a few seconds,as observed in Figure 2. The frequency of a TV watch-ing duration increases from 1 to 4 seconds, then dropsvery quickly. We have looked deeper into the 1 to 4seconds area of the graph and we have identified thatit is compound of two different issues: one is the vari-ation in the channel switching time provided by theIPTV system studied and the time required for peo-ple to quickly recognize a program. However, the exactimpact of these issues need further investigation.

1 sec 4 sec 1 min 10 min 1 hour

102

104

106

Duration of TV watching

Fre

quen

cy

Figure 4: Duration of time watching each channel

4

Table 1: Breakdown of popularity across genre (probability of a viewer watching each genre)genre free mixed kids docu local cine sports music news audio the rest total

viewing prob. 38.6% 21.5% 12.5% 6.6% 4.9% 3.9% 3.8% 2.3% 1.0% 0.3% 4.6% 100%num. channels 6 19 7 12 17 6 8 11 13 15 36 150

∗Genre categorized “the rest” includes pay-per-view, satellite, and promotional channels.

The graph after the 4-second mark follows a power-law distribution [10, 21]. This reflects an interestingtendency of users to continue watching the same chan-nel: the probability of continuing to watch a channelincreases proportionally with the length of time alreadyspent watching that channel. This is in contrast tothe exponential distribution, which has generally repre-sented human activity in environments, such as wirelessaccess [8] and telephone call [26] durations, and reflectsthe memoryless property that future activity is inde-pendent of past activity. According to Vazqueza [30],the power-law pattern in humans could be due to peoplehaving some perception of their past activity rate (i.e.,not memoryless) and thereby reacting by acceleratingor reducing their activity intensity [11, 14].

4.1.3 Attention Span

When watching television, people browse through a setof on-air TV programs until they find something inter-esting. The behavior of channel selection involves thefollowing two steps: (a) browsing of content to decidewhether to continue or stop streaming, and (b) switch-ing through multiple channels for repeated browsing,until a desired channel is found. In Figure 2, we observethat over 60% of user action is related to channel surf-ing. To understand in detail how long a single sampletakes, we summarize the characteristic channel holdingtimes for a few channel genres in Table 2.

Table 2: The time spent in surfing and viewinga single TV program grouped by channel genre

music free docuperc. news kids sports mixed cine50th 6 sec 6 sec 7 sec 9 sec 9-11 sec90th 3 min 10 min 2-3 min 6-7 min 5-6 min95th 9 min 21 min 5-8 min 17-18 min 15-17 min

The 50th percentile values range from 6 to 11 seconds,indicating that viewers watched only for a very shortwhile and changed the channel. The median durationis slightly longer for kids, free and mixed. The 90 and95th percentile values reflect the viewing durations andshow larger variations ranging from 5 to 21 minutes. Atthe 90th and above percentiles, the viewing durations ofmusic and sports are much shorter than those of kids,free, mixed, and cine – or conversely, the attentionspan is much shorter when watching music or sports.

Table 1 summarizes the percentage of time spent oneach genre, over the trace period. free and mixed gen-res account for over 50% of TV viewing instances. [2]reports that these genres similarly account for over 60%of TV viewing in Canada. Regional content is also pop-ular, considering that each region has one or two local

channels. Kids and docu programs are also popular.There are significant differences in our findings of chan-nel genre popularity and those reported by samplingstatistics for the US population [25]. For instance, [25]reports kids and docu channels to almost double thepercentage in viewing share we find in our analysis.While this could well be a trend in different social be-haviors in the analyzed populations, it could also beattributed to a change in viewing behavior when house-holds are under the scrutiny of a tracking device.

4.2 Channel Popularity and DynamicsThe current IPTV system continuously sends out mul-ticast streams of all channels throughout the destina-tions in the backbone (i.e., DSLAMs). However, suchfixed distribution may no longer be possible if and whenthe number of channels increases. In certain cases, thebackbone needs more bandwidth to deliver channels inmulticast rather than in unicast. One strategy to min-imize the network cost in the backbone is to use multi-cast for some channels, while carrying the rest as unicast(or P2P distributed fashion [13]). Accurate measure-ment of the channel popularity is critical in designingsuch systems. In this section, we study channel popu-larity from three different angles. First we analyze itsdistribution to understand how user interests are spreadacross channels. We then investigate how fast channelpopularity changes over time, and what causes such dy-namics.

4.2.1 Popularity Distribution

We start with the channel popularity distribution. Fig-ure 5(a) shows the aggregate fraction of viewers ac-counting to the least x-th popular channels. The hori-zontal axis represents the ranking of channels normal-ized between 0 and 100, from the most popular to theleast popular. The top 10% of channels account fornearly 80% of viewers, which reflects the Pareto prin-cipal or 80-20 rule. This is consistent across differenttimes of the day. Figure 5(b) shows the popularity dis-tribution as a function of the channel ranking againstthe number of viewers, where the unpopular channelsare put at the tail. The plot shows a straight line

5

0 20 40 60 80 1000

20

40

60

80

100

Normalized channel rankAgg

rega

te v

iew

er s

hare

(%

)

8AM3PM10PM

(a) Pareto principal

100

101

102

100

102

104

Channel rank

Num

ber

of v

iew

ers

10PM3PM8AM

(b) Empirical popularity of channels

Figure 5: Channel popularity distribution

waist for top ranked channels (below 30) and a trun-cated tail. This indicates that popularity is Zipf-likefor top channels and it decays fast for non-popular ones.Other live video streaming systems have showed similartrends [19, 29, 33].

In our IPTV system, some channels also fail to at-tract enough viewers compared to a normal Zipf dis-tribution. The set of channels that ranked below 100are (a) individual channels that users had to pay in “ala carte” billing such as pay-per-view and certain doc-umentary and sports channels, (b) foreign news, e.g.,Russia, France, and Al Jazeera, which is spoken in dif-ferent language, (c) local channels of small regions,and (d) audio and reality shows.

4.2.2 Popularity Dynamics

We next examine popularity dynamics, in particular,how the channel ranks change over time. We examinethe channel ranks for every 30 minute of a representativemonth period. At each time interval, we identify thegroups of channels that make up the top 10, 30, and 50ranks. We break ties in ranking by comparing the totalviewing time of channels. We ran the analysis over thetrace period and did not find differences between days.In Figure 6(a), we show the percentage of viewer sharefor the top ranked channels for a representative day.The average viewer shares are similar to that shown inFigure 5(a). However, the graph shows significant fluc-tuations across the day. Interestingly, popularity of topchannels is reinforced around 8AM, 3PM, and 10PM.We conjecture this is due to the role of TV viewing asa “family” activity, where what to watch is no longera single person’s decision but is shared amongst familymembers. (See [24] for a discussion of how group de-cisions are taken when watching television.) The userinterests are more widely spread across channels duringother times of the day.

Our next question is how frequent channel rankschange. Compared to the analysis on stability of pop-ularity distribution in Figure 6(a), rank changes indi-

cate the stability of popularity for individual channels.To measure the variations in channel ranks, we mon-itor the channel ranks for every 5-minute of the traceperiod and calculate dissimilarity coefficient for degreeranks ξ = 1− ρ2, where ρ is the Spearman’s rank-ordercorrelation between channels of two time intervals3. ξvaries between 0 and 1 and measures the proportion ofvariance in degree ranks that cannot be predicted fromthe initial set of ranks.

We compare the dissimilarity coefficient betweenchannel rank snapshots of the first 5-minute time in-terval of Day 1 and the subsequent time intervals. Weran the analysis over the trace period and did not finddifferences between days. In Figure 6(b), we show theresult for two representative days. Typically, a dissimi-larity coefficient greater than 0.1 is considered to havesubstantial change in ranks.

Figure 6(b) shows periodic patterns over the two dayperiod, indicating daily rank changes are repetitive. Weexpect scheduled programs to cause such periodic pat-tern. The dissimilarity coefficients increase at low usagehours, above 0.1. This indicates substantial degree ofrank changes during those hours, supporting our findingin Figure 6(a). A detailed look into the set of channelsthat compose the top popular list shows that 36, 63 and90 distinct channels made the top 10, 30, and 50 ranksover the trace period. Only 1 channel consistently madethe top 10 list. Our results suggest that while channelpopularity distribution is skewed at any given moment,the set of channels that compose the top popular listchanges dynamically over the course of a day.

3Spearman’s rank correlation compares the two set of val-ues by converting each variable to ranks [28]. Kossinets andWatts, in their work to quantify the stability of node degreein social networks [22], also used Spearman rank-order corre-lation as dissimilarity measure. In our case, the correlationis calculated,

ρ = 1 −6

P

d2

i

n(n2 − 1), (1)

where di is the difference between ranks of the same channelat two time intervals, and n is the number of channels.

6

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM0

20

40

60

80

100

Time of day

Vie

wer

sha

re (

%)

Top 50 channelsTop 30Top 10

(a) Viewer share of popular channels

12AM 6AM 12PM 6PM 12AM 6AM 12PM 6PM 12AM0

0.05

0.1

0.15

0.2

0.25

Time (2 days)

Dis

sim

ilarit

y co

effic

ient

Day 1 Day 2

(b) Degree of change in channel ranks over two-day period

Figure 6: Channel popularity dynamics

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM0

1

2

3

4

5

6

7x 10

4

Time of day

Num

ber

of v

iew

ers free

mixed

kidslocal

the rest

(a) The number of viewers per genre

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM0

10

20

30

40

50

60

70

Time of day

Vie

wer

sha

re (

%)

freemixedkidsthe rest

(b) The percentage of viewer share per genre

Figure 7: Breakdown of viewers across genre over a representative day (Monday)

One possible cause of the short-term focus on partic-ular channels but long-term diversity of popular chan-nels is that there are certain genres of programs thatare periodically popular during certain times of the dayor week. Also, one-time events may suddenly becomepopular for a short period of time, for instance, theAcademy Awards. To examine the temporal correlationbetween popularity and genre, we show the aggregatenumber of viewers across content genres in Figure 7(a).We count the number of distinct viewers for every 10-minute time interval of a representative day. As ex-pected, free content is the most popular throughoutthe day, except for the mid-day and around dawn whenkids and mixed genres are the number one line up,respectively. While all other genres show diurnal pat-tern, the number of viewers for kids genre are relativelysteady from early morning to late night. Figure 7(b)further shows the percentage of viewer share for free,mixed, and kids, and the remaining genres. The per-centage of viewer share adds up to 100% at any momentand the top popular genre changes over a day.

4.3 Geographical LocalityWe now identify the content locality – the degree towhich similar TV programs are watched by users ingeographical proximity. We obtained the informationabout geographical location of users from the IPTV

provider. Naturally, we expect people to show moreaffinity to the local channels belonging to each region.However, it is not clear that the rest of channels willhave a similar affinity across different geographical loca-tions. We compare the spatial locality of users’ interestsat different levels of granularity (regions and DSLAMs)and over genres and channels.

4.3.1 Locality Across Regions

We first measure which content genres are popularacross regions. Figure 8 shows the relative popular-ity of genre (based on the total time watched) for the12 regions across a month period. We observe that therelative ranking of genre is similar across regions – thecorrelation coefficient ranged from 0.97 to 0.99 betweenany two pairs of regions. However, the probability ofwatching regional content varied widely from 1% to 9%.This shows that preference towards local channels arestronger for some regions versus others. Overall, we findthat the spatial variations in preference in genre acrossregions are up to 20%.

4.3.2 Locality Across DSLAMs

To analyze geographical locality in a smaller scale,we investigate how two users in two different DSLAMsare similar in their viewing patterns. We focus on threeexample channels: channel A, a top ranked free chan-

7

1 10 100 6700

10

20

30

40

50

DSLAM IDs sorted by viewer share

Cha

nnel

vie

wed

tim

e (%

)

(a) Channel A

1 10 100 6700

5

10

15

20

25

DSLAM IDs sorted by viewer share

Cha

nnel

vie

wed

tim

e (%

)

(b) Channel B

1 10 100 6700

5

10

15

DSLAM IDs sorted by viewer share

Cha

nnel

vie

wed

tim

e (%

)

(c) Channel C

Figure 9: Comparison of locality across 670 DSLAMs based of three example channels

1 2 3 4 5 6 7 8 9 10 11 120

20

40

60

80

100

kidscinedoculocalrest

free

kidscinedoculocal

rest

Region ID sorted by the number of users

Bre

akdo

wn

of g

enre

pop

ular

ity

free

mixedmixed

Figure 8: Popularity of genre across regions

nel; a kids channel B, which is ranked 10th; and a docu

channel C, which is ranked below 30. We calculate, forall 670 DSLAMs, the ratio of the time spent on viewingof the corresponding channel against the total viewingtime of all online users within each DSLAM. Figure 9shows the result for a 9PM to 11PM trace of a represen-tative day. DSLAMs are sorted in each figure based onthe aggregate viewer share of the corresponding chan-nel. To easily distinguish the case when no single userwatched the channel, we only plot for those DSLAMswith at least one viewer of the corresponding channel.

Figure 9 shows that the aggregate viewer share ofchannel A varies largely (10-42%) across DSLAMs. Thismeans that 42% of total viewing time of the users wasspent on channel A for one DSLAM, while for someDSLAMs only 10% of the total viewing time of onlineusers was spent on channel A. The mean fraction oftime spent is 24% for channel A. Channel B shows lessvariations (0-25%) across all DSLAMs and the meanpercentage of time spent on the channel is 6%, which ismuch smaller than that of channel A. Finally, only 134DSLAMs show at least one viewer for channel C duringthe trace period. The mean time spent on the channelis much lower, 0.2%.

4.4 SummaryIn this section, we have characterized the distribution ofpopularity of channels and studied the popularity evolu-tion over time. Despite the fact that at any given point,

the popularity is highly skewed over a few channels,those few channels change over time, making the systemhighly dynamic. We have also found that though cer-tain channels are popular at any given time, the set ofchannels that are popular across regions and DSLAMsis not always the same. We have in fact found the vari-ations up to 20% in the set of channels that people arewatching across different regions or DSLAMs within thecountry.

5. HOW PEOPLE FIND CONTENTIn this section, we analyze channel switching behav-iors to gain understanding of (a) what affects channelchanges, (b) how people find interesting programs towatch, and (c) how people join and leave a particularchannel.

5.1 Factors that Affect Channel ChangesTable 2 shows that attention span of a user is typicallygenre-dependent and ranges from 5 to 21 minutes; genreclearly affects the likelihood and frequency of channelchanges. In this section, we focus on other potentialfactors of channel changes: the time of day, programpopularity, and program length. For each of these fac-tors, we calculate the probability that a given channelchange is due to surfing compared to viewing, or whatwe define as the surfing probability4, to infer the user’sreason for channel switching. A surfing probability of0 indicates that after the channel change, the user willstay and view. While the probability of 1 means thatchannel change of a user will be followed by anotherchannel change, indicating the user is surfing.

We first investigate how the surfing probability re-lates to the time of day. Figure 10 shows the minimum,the median, the average, and the maximum values ofsurfing probability for every 10 minutes of a month-long period. The median and the average values rangefrom 0.65 to 0.90. The surfing probability decreases4Let J be the set of channel join events during a given timeinterval, then we calculate the surfing probability as,

|surfing events in J |

|surfing or viewing events in J |. (2)

8

around 6-7AM, 3PM, and 9PM, which are the peakIPTV usage hours. In conjunction with our finding onthe reinforced popularity of top ranked channels at peaktimes, the decrease in surfing probability at these hoursindicates that a larger pool of concurrent viewers aremore predictable and stable in their viewing patterns.The daily scheduled programs likely cause them to surfless during peak usage hours. The lowest point of surf-ing probability comes earlier than 8AM in the morning.We assume this phenomenon is related to long programlength: three top ranked channels had a 3-hour longprogram from 6AM to 9AM. In contrast, we see in-crease in surfing probability in late evening and mid-night, when the top popular programs are shorter inlengths (30 minutes to 1 hour) and when we expect alot of people finish up some activity and check out whatis on TV.

12PM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM0.4

0.5

0.6

0.7

0.8

0.9

Time of day

Pro

babi

lity

of s

urfin

g

Min−MaxMedianAverage

Figure 10: Time of day and channel changes

Next, we investigate the influence of program pop-ularity on channel changes. We compare the channelranks and the surfing probability for every 10 minutesof a month period. Figure 11 shows the 10th percentile,the median, the average, and the 90th percentile valuesof surfing probability against the program ranks. Thesurfing probability drops by 20% for 10% of the mostpopular programs. This indicates that viewers are morelikely to stay tuned once they join a popular program.The correlation between the surfing probability and theprogram ranks is less obvious for medium to low rankedprograms. For very unpopular programs (ranks beyond100), we see increase in surfing probability. That is,users visiting unpopular channels are likely to switchchannels immediately.

We now look at the program length as the third po-tential factor on channel changes. Figure 12 shows thesurfing probability of channel join events grouped bythe absolute time passed since the beginning of the pro-gram. We show the 10th percentile, the median, theaverage, and the 90th percentile values over every 5-minute time interval. The graph shows that the surfingprobability is insensitive to the time offset of programswithin which users join. When we repeat our analysisover other months of the trace, we consistently observea slight increase of surfing probability at 15-20, and 30-

0 20 40 60 80 100 120 1400

0.2

0.4

0.6

0.8

1

Program ranks

Pro

babi

lity

of s

urfin

g

← People surf less when watching popular programs

10−90 percentileMedianAverage

Figure 11: Program popularity and channel

changes

35 minute offsets. This may be due to users changingchannels during commercial breaks. The graph is noisyafter an offset of 60 minutes because there are fewerdata points.

0 30 60 90 120 150 180 2100

0.2

0.4

0.6

0.8

1

Absolute time offset within a program (minutes)

Pro

babi

lity

of s

urfin

g

10−90 percentileMedianAverage

Figure 12: Program length and channel changes

So far, we have used the viewing period of 1 minute todemarcate surfing instances from viewing instances. Asgiven in Equation (2), the surfing probability is sensi-tive to the choice of viewing period. For validation, werepeated the three types of analysis in Figures 10, 11and 12 over different viewing periods: 10 seconds, 30seconds, 2 minutes, and 5 minutes. We consistentlyfound the same correlation between the surfing proba-bility and the three potential factors.

5.2 Switching From One Channel To AnotherWith a TV remote control, people switch channels upand down easily in a sequential fashion. We call suchchannel changes linear. In IPTV, people may choosea different channel change strategy because channelswitching time takes longer (typically within 1 second,compared to 200 milliseconds in conventional TV) andthere are simply too many channels. Electrotonic pro-gram guide (EPG) is a popular way to change channelsin IPTV. Using EPG, viewers can browse the broadcast-ing schedule overlayed on the screen and jump to arbi-trary channels. Users also can directly input channel

9

Channel IDs sorted

Cha

nnel

IDs

sort

ed

free

mixedcine

kidsmixed

docu

sports

music

newssatellite

ppv

audio

20 40 60 80 100

10

20

30

40

50

60

70

80

90

100

110

(a) Zapping probability from channel x to channel y

Channel IDs sorted

Cha

nnel

IDs

sort

ed

free

mixedcine

kidsmixed

docu

sports

music

newssatellite

ppv

audio

20 40 60 80 100

10

20

30

40

50

60

70

80

90

100

110

(b) Viewing probability from channel x to channel y

Figure 13: Linear and non-linear channel change patterns for zapping and viewing events from channelx to channel y

numbers through their TV remote controls and jumpto any channel. The sequence of channels a user visitsin such manner no longer needs to be sequential. Wecall the channel changes between a pair of channels thatare not sequentially numbered non-linear.

To understand the collective switching behaviors ofall users, we examine the linear and non-linear switch-ing patterns. Using a month-long trace, we calculatethe normalized average probability of channel changesbetween every pair of channels and visualize them inFigure 13(a). While the actual channel numbers rangefrom 0 to 265, we remove gaps between channel num-bers and reassign them sequentially in the matrix. Acolor pixel at (x, y) represents the probability of channelswitching from channel x in the horizontal axis to chan-nel y in the vertical axis. Darker pixel reflects higherprobability. Since most channel changes are concen-trated on a few pairs of channels, we use the log valueof the probability to visualize both frequent and infre-quent events. Therefore, any slight difference in thebrightness of pixel is noteworthy.

Figure 13(a) shows several interesting channel switch-ing habits. First, over 60% of channel changes are lin-ear. The upward linear switches (on line y = x + 1)account for 43% and the downward linear switches(y = x − 1) account for 19%. Because channel changeshappen between two different channels, the switchingprobability on line y = x is zero. Second, certain genresshow a distinctive pattern of non-linear channel changeswithin the genre, e.g., free, sports, and kids. Thisreflects that users are interested in the set of channels ofsimilar topic. For example, a user tuning into one sportschannel is likely to check out other sports channels,

compared to movies or documentaries. In total, 76%of channel changes are within the same genre group,where 58% of switches are linear and 18% non-linear.

Third, the pattern of linear channel changes contin-ues in less popular channels like music, satellite, andaudio. This indicates that linear channel changes, evenwith its inconvenience in IPTV compared to conven-tional TV, remain a characteristic behavior of users. Fi-nally, 18% of channel changes are non-linear across dif-ferent genres. Based on our findings above, we concludethat an average user switches channels non-linearlyfrom one genre to another, then switches linearly withinthe genre group.

Next, we examine the influence of channel change pat-terns on viewing. Using the same trace, Figure 13(b)visualizes the probability of channel changes that led toviewing events, from channel x in the horizontal axisto channel y in the vertical axis. Intermediate surfingor away events are excluded in the analysis5. As inFigure 13(a), a darker color represents logarithmicallyhigher viewing probability. Compared to Figure 13(a),one striking difference is the consecutive viewing of thesame channel on line y = x, which now accounts for 17%of all viewing instances. This reflects users’ tendency toreturn to the same viewed channel for continued view-ing, even after surfing. For example, viewers browsethrough other on-going channels during commercialsand rejoin the viewed channel. This may lead to miss-

5The sequence of two viewing events may not be relevantif they are separated by away events. The results in thissection are consistent even when we ignore viewing eventsseparated by away events.

10

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM

2000

4000

6000

8000

Time

# us

ers

/ min

(a) User arrival rate of channel A

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM

2000

4000

6000

8000

Time

# us

ers

/ min

(b) User departure rate of channel A

Figure 14: User arrival and departure evolution of a top popular free channel

ing some part of the programs, especially when usersre-join few minutes late after the commercial break.

Another distinctive difference is the extent of non-linear viewing. Non-linear viewing patterns account for67% of viewing instances, which is two times higherthan that of switching probability. Non-linear viewingis increased for both inter-genre (40%) and intra-genre(27%) channel changes. The total intra-genre viewinginstances account for 60%, indicating that viewers areinclined to watch programs of a similar topic. Within agiven genre, some channels are viewed much less thanothers. This is due to the popularity differences of chan-nels and the fact that some users are only subscribingto some channels (e.g., basic, family, a la carte).

In summary, viewers tend to continue watching thesame channel even after switching for some time andwith high probability. This effect is called “inheritance”and is a well-established fact in traditional audience re-search (see [32], for instance).

5.3 User Arrival and Departure PatternsTo understand the dynamics of user churn in the sys-tem, we analyze users’ arrival and departure patternsof individual channels. We first demonstrate the arrivaland departure rates for an example channel. We theninvestigate the inter-arrival and inter-departure timesto infer the distribution of those processes.

5.3.1 Arrival and Departure Rates

Minute-by-minute user arrival and departure patternsdemonstrate how users join and leave a particular chan-nel. We discuss our findings on user arrival and depar-ture processes using an example channel A, a top rankedfree channel. Figure 14 shows the arrival and depar-ture rates of users in channel A, over a representativeday. The arrival and departure rates are similar on av-erage. However, the latter ranges widely from 18 to7,720 users per minute, while the former ranges from18 to 4,172.

We make several observations. First, the arrival anddeparture rates vary over the day. A detailed analysisreveals that the variation is related to the viewer base.Figure 15 shows the number of concurrent viewers who

watched channel A for longer than 1 minute in every 10minutes of the day. The maximum values of user arrivaland departure rates in Figure 14 follow the viewer basewell. Our finding agrees with observation in other videosystems [19, 29, 33].

12AM 3AM 6AM 9AM 12PM 3PM 6PM 9PM 12AM0

5,000

10,000

15,000

20,000

Time

# vi

ewer

s

Figure 15: Number of viewers for channel A

Second, user departure patterns show consecutivespikes as in Figure 14(b), where the departure rate sud-denly increases up to 5 times. These spikes are due tomany users leaving the channel simultaneously at theend of programs or at advertisements. Hei et al. [19]also observed a batch departure pattern in P2P IPTVsystem for a popular movie channel. However, theydid not find such pattern in regular TV channels. OurIPTV system show spikes in departure patterns for pop-ular channels across all genres. The difference betweengenres is the interval between spikes: cine channels hada gap of 2 hours, while other channels showed shorter in-tervals between spikes. Some of the non-popular chan-nels with few tens of concurrent users also show thebatch departure patterns.

Third, the user arrival is much less time-correlatedthan the departure. Unlike Figure 14(b), the arrivalpatterns show much less variations. This indicates thatthe number of users watching a channel ramps up grad-ually over time, while it drops sharply at batch depar-ture processes.

5.3.2 Inter Arrival and Departure Times

Finally, we look at the inter arrival and departure timesto identify the arrival and departure processes. Fig-ure 16 shows the cumulative distribution of inter arrivaland departure times for the popular channel A, whichis the top popular channel. The granularity of chan-

11

nel changes are 1 second in our trace and we assumearrivals are uniformly distributed within each second.The median inter arrival time is 0.07 seconds, indicat-ing that 1/0.07 = 14 users join channel A per second.The median inter departure time is also 0.07 seconds;users leave and join channel A at the same rate. The ar-rival rate varies over time and the arrival process is notstationary over the course of a day. The maximum in-ter arrival time is 14 seconds, which corresponds to thelow usage hours. The maximum inter departure time is700 seconds. While we do not show here, we find thatboth arrival and departure processes at short timescales(less than 1 millisecond) are exponentially distributed.This observation is consistent with previous studies onarrival processes of live streaming servers [7, 29, 31].

0.001 0.01 0.1 1 150

0.5

0.91

Inter arrival times (second)

CD

F

(a) User inter arrival times of channel A

0.001 0.01 0.1 1 10 100 7000

0.5

0.91

Inter departure times (second)

CD

F

(b) User inter departure times of channel A

Figure 16: Inter arrival and departure time dis-tribution

5.4 SummaryIn this section, we have identified how people find con-tent. Understanding this mechanism is important be-cause any user activity to search for content, imposessome load in the network. Surprisingly, we find thatmost of the channel changes do not lead to viewing ac-tivity, which highlights the fact that the network in-frastructure could be overloaded with superfluous shortchannel surfing events. Our finding has major impli-cations to IPTV networks, as opposed to traditionalbroadcast TV networks (e.g., cable and satellite TV).This is because in traditional broadcast networks, allchannels are delivered to the end users and surfing ac-tivities of users do not create additional load on thenetwork.

We have also characterized the arrival and departurepatterns of users, where we have found time-correlatedarrival and batch-like departure processes. The dynam-ics of user churn puts significant challenges to the designof peer-to-peer (P2P)-based IPTV systems as they needto deal with drastic changes of the viewer population ina short period of time.

6. IMPLICATIONSOur observations about IPTV workloads can help re-searchers to understand the user and the content char-acteristics of TV and video streaming systems. Ourfindings offer useful insights for the designers of exist-ing and future video applications. For example:

• Channel Selection Process: Most channel changesof a user are related to channel surfing and are short-intervaled. Therefore assisting users in channel selec-tion is crucial to the quality of the experience. In cur-rent IPTV and Internet video systems, electronic pro-gram guides (EPG) and metadata about channels (suchas tags and descriptions) provide users with useful addi-tional information about content. It is also importantto make new content quickly available at the time ofchannel switching. This is more challenging for high-bandwidth video and TV streaming, especially in peer-to-peer (P2P) distribution systems. For instance, thestate of the art P2P Internet TV service, Joost, takes upto 5 seconds to switch channels. Infrastructure-basedTelco’s IPTV systems guarantee a switching delay ofless than 1 second. However, this is not satisfactory asa 100-200 ms switching delay is considered “instanta-neous” by viewers.

Our results about linear channel changes suggest thatproviding the channel numbers sorted for each user byeither channel popularity for that user or dynamicallybased on current program popularity might reduce un-necessary channel changes. Our findings on the peri-odic popularity dynamics further suggest that a moredynamic re-numbering following time of day trends willbe needed.

• Popularity and Program Ratings: An accuratemeasure of popularity is useful for the TV viewers, con-tent producers, and distributors. The popular Neilsenratings are usually aggregated over a long period of timeand delayed few hours to a day. With IPTV, we can getthe instantaneous popularity and monitor its evolutionvery closely. Having access to such realtime informa-tion is useful for TV viewers in deciding what programto watch.

• Rewind and On-Demand Features: Traditionaland the first generation IPTV systems are live. NowTV is transforming itself to interactive on-demand ser-vice and users are allowed to rewind back to any pointback in time and watch TV at the time they want. Our

12

findings on user arrival and departure patterns demon-strate that users’ arrivals to programs are less corre-lated than departures. Also users take a temporaryleave during commercials and come back late, missingout on the beginning part of the programs they watch.These have significant implications on the design of apotential peer-assisted IPTV rewind service. In our ear-lier work [13], we studied the potential for such peer-assisted rewind service where users joining a channelcan rewind back to the beginning of the on-going pro-gram.

7. RELATED WORKUnderstanding TV viewing habits in the old days reliedon specialized monitoring boxes or phone surveys. Dueto the difficulty of conducting a large-scale survey, themonitoring was limited to a subset of the TV viewingpopulation. A representative example is the NielsenMedia Research [4] that provides a wide set of statisticson TV viewing and program ratings.

Viewing behaviors in Internet streaming have re-ceived recent attention. Sripanidkulchai et al. [29] stud-ied the Internet radio and video streaming workload of5,000 users in Akamai, the largest content distributionnetwork. They observed a diurnal pattern in traffic anda Zipf-like popularity of live content. They found aheavy-tailed session duration and explained that shortmedian session durations (less than 10 seconds) arelikely due to channel surfing events. Our results areconsistent with these earlier findings, but they are basedon a specific IPTV system and the number of users ismore than two orders of magnitude larger.

There are peer-to-peer (P2P)-based Internet videostreaming systems such as Joost, Zatoo, and Microsoft’sLiveStation, that provide TV like viewing experience.Hei et al. [19] studied the packet traces from PPLive,a P2P based IPTV system, and showed that peers joinand leave at a higher rate at peak times. They ob-served a periodic (every 2 hour) batch user departurepattern for a popular movie channel. We also observedbatch like departure pattern, however, at a smaller timescales (tens of minutes). This difference may be due tothe channel switching delay of the two systems. Unlikeour IPTV system, P2P IPTV has long start-up delaysthat range from tens of seconds to several minutes [19].Due to the long playback lags users may behave some-what differently from other TV viewing environments,for instance, users may be reluctant to switch channelsduring commercials. Other studies similarly analyzedthe workloads and traffic characteristic of P2P IPTVsystems [27, 34]. These results complement ours; whileour system is infrastructure-based, understanding bothsystems provide a comprehensive view of IPTV work-loads. In our earlier work [13], we looked into the po-tential for peer-assisted IPTV system.

Many previous studies have measured network prop-erties of video-on-demand systems [15, 16, 19, 20, 33].Cherkasova and Gupta [16] analyzed the enterprise me-dia server workloads from Hewlett-Packard Corpora-tion; Yu et al. [33] measured a commercial video sys-tem in China, PowerInfo; and Cheng et al. [15] ana-lyzed the workloads of their high-bandwidth P2P-basedVoD system. They discuss the popularity distributionof videos, viewing session characteristics, and their im-plications on designing efficient distribution architec-tures. Although our paper uses similar measurementtechniques, the network environment we study is differ-ent as the viewing in live IPTV is not on-demand, butlargely affected by the broadcaster’s schedule, and oursis the first study is to use large-scale real traces from acommercial telco’s IPTV service.

8. SUMMARY AND OUTLOOKTelevision has long been a dominant and pervasive massmedia and its viewing behavior has many important im-plications on the video distribution networks and sys-tems. In this paper, we presented the first large-scalemeasurement study of a telco’s IPTV system. We char-acterized several properties of viewing sessions, chan-nel popularity dynamics, content locality, and channelswitching behaviors. Our findings have important im-plications to the design of the open Internet TV distri-bution systems like Joost, as they try to replicate IPTVon the wider Internet and will face the same challengesof IPTV systems.

There are several directions that we wish to pursueas future work. The first is to understand the implica-tions of the workloads on the system design. Our traceshowed that over 60% of a user’s channel changes arerelated to channel surfing. We are interested in design-ing a system that can reduce the number of channelsusers need to sample before finding something that in-terests them. Second, we would like to study the pro-files of individual users. While we limited our study tounderstanding the behavior of an average populace, fu-ture studies can focus on classifying users into differentprofiles. For example, what are the various user pro-files based on the time they watch TV, channels theywatch, and the program genre they are interested in?Answering these questions are promising in the contextof IPTV recommendation systems [9]. Finally, it wouldbe very useful to develop a model that captures userswitching behaviors and content popularity dynamics.The designers of future streaming systems can utilizesuch realistic models to test and assess their systems.

9. REFERENCES[1] Australian Film Commission. Viewing Patterns:

Hours of Viewing by Age Group, 2004.http://www.afc.gov.au/gtp/wptvviewxage.html.

13

[2] Statistics Canada. Television Viewing Data Bank,2002.

[3] IMS Research. A Global Market Analysis, 2008Edition.

[4] The Nielsen Company.http://www.nielsenmedia.com.

[5] The Nielsen Company. Digital Video Recorders Growin Popularity, 2007.

[6] Office For National Statistics, United Kingdom. TimeSpent on Selected Free Time Activities of Full TimeWorkers, 2001.

[7] K. C. Almeroth and M. H. Ammar. Collecting andModeling the Join/Leave Behavior of Multicast GroupMembers in the MBone. In International Symposiumon High Performance Distributed Computing (HPDC),1996.

[8] H. Anderson. Fixed Broadband Wireless SystemDesign. Wiley, 2003.

[9] L. Ardissono, C. Gena, P. Torasso, F. Bellifemine,A. Difino, and B. Negro. Personalized DigitalTelevision. Targeting Programs to Individual Users,chapter User Modeling and RecommendationTechniques for Personalized Electronic ProgramGuides. Kluwer, 2004.

[10] A.-L. Barabasi and R. Albert. Emergence of Scaling inRandom Networks. Science, 1999.

[11] D. Brockmann, L. Hufnagel, and T. Geisel. TheScaling Laws of Human Travel. Nature, 2005.

[12] M. Cha, G. Choudhury, J. Yates, A. Shaikh, andS. Moon. Case Study: Resilient Backbone NetworkDesign for IPTV Services. In Workshop on IPTVServices over WWW, 2006.

[13] M. Cha, P. Rodriguez, S. Moon, and J. Crowcroft. OnNext-Generation Telco-Managed P2P TVArchitectures. In IPTPS, 2008.

[14] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass,and J. Scott. Impact of Human Mobility on theDesign of Opportunistic Forwarding Algorithms. InIEEE INFOCOM, 2006.

[15] B. Cheng, L. Stein, H. Jin, and Z. Zheng. TowardsCinematic Internet Video-On-Demand. In ACMEurosys, 2008.

[16] L. Cherkasova and M. Gupta. Analysis of EnterpriseMedia Server Workloads: Access Patterns, Locality,Content Evolution, and Rates of Change. IEEE/ACMToN, 2004.

[17] S. E. Deering and D. R. Cheriton. Multicast Routingin Datagram Internetworks and Extended LANs.ACM TOCS, 1990.

[18] T. Gitlin. Inside Prime Time. University of CaliforniaPress, 2000.

[19] X. Hei, C. Liang, J. Liang, Y. Liu, and K. Ross. AMeasurement Study of a Large-Scale P2P IPTVSystem. IEEE Trans. on Multimedia, 2007.

[20] C. Huang, J. Li, and K. W. Ross. Can Internet VoDbe Profitable? In ACM SIGCOMM, 2007.

[21] Y. Ijiri and H. A. Simon. Skew Distributions and theSize of Business Firms. ElsevierNorth Holland, NewYork, 1997.

[22] G. Kossinets and D. J. Watts. Empirical Analysis ofan Evolving Social Network. Science, 2006.

[23] M. Levy and S. Windhal. Media GratificationsResearch, chapter The Concept of Audience Activity.Sage, 1985.

[24] J. Masthoff. Group Modeling: Selecting a Sequence ofTelevision Items to Suit a Group of Viewers. UserModeling and User-Adapted Interaction, 2004.

[25] S. Pedder. Power in Your Hand. The Economist, April2002.

[26] P. Reynolds. Call Center Staffing. The Call CenterSchool Press, 2003.

[27] T. Silverston and O. Fourmaux. Measuring P2P IPTVSystems. In NOSSDAV, 2007.

[28] C. Spearman. The Proof and Measurement ofAssociation Between Two Things. American Journalof Psychology, 1904.

[29] K. Sripanidkulchai, B. Maggs, and H. Zhang. AnAnalysis of Live Streaming Workloads on the Internet.In ACM IMC, 2004.

[30] A. Vazqueza. Impact of Memory on HumanDynamics. Physica A, 2006.

[31] E. Veloso, V. Almeida, W. Meira, A. Bestavros, andS. Jin. A Hierarchical Characterization of a LiveStreaming Media Workload. In IMW, 2002.

[32] J. G. Webster. Audience Flow Past and Present:Television Inheritance Effects Reconsidered. Journalof Broadcasting & Electronic Media, 2006.

[33] H. Yu, D. Zheng, B. Y. Zhao, and W. Zheng.Understanding User Behavior in Large-ScaleVideo-on-demand Systems. In ACM Eurosys, 2006.

[34] X. Zhang, J. Liu, B. Li, and T.-S. P. Yum.CoolStreaming/ DONet: A Data-driven OverlayNetwork for Peer-to-Peer Live Media Streaming. InIEEE INFOCOM, 2005.

14