The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)
SERVICE PROVISIONING AND TERMINAL COOPERATION IN USER-DEPLOYED NETWORKS Nicolas Debernardi, Markus J¨ager, Jean-Christophe Laneri, Malek Sraj and Pietro Lungaro Wireless@KTH, The Royal Institute of Technology Electrum 418, S-164 40 Kista, Sweden A BSTRACT Exploiting high-speed user-deployed Access Points (APs) in public infrastructures has been proposed as one of the candidate solutions for reducing the cost of future radio access. A disadvantage is that, due to lack of deployment coordination, these types of network are likely to provide only partial coverage. For this reason, the set of services that could be successfully provided in these networks might be rather limited. In this paper we present a novel framework for modeling non-interactive “infotainment” services with different degrees of “time criticality”, and utilize it for investigating the user service perception in “spotty” coverage networks. In order to hide the infrastructure sparsity to the end users, the terminals are assumed equipped with software agents that can “opportunistically” pre-fetch information, on behalf of their users. This can be done when in contact with an AP, or, through a peer-to-peer file exchange with other terminal agents that have previously accessed the same information. In this paper, for different service types, the relationship between AP density and service perception is investigated. Furthermore, the impact of peerto-peer terminal cooperation is evaluated in respect to both infrastructure requirements and content access delay. The results show that already with moderate AP densities, user-deployed networks deliver acceptable service perception, especially for services with low time criticality. Furthermore, whenever a “critical” mass of users shares common interests, the adoption of a peer-to-peer exchange strategy brings the significant gains of infrastructure reduction and/or improved service perception. I.
networks of confederated APs [5] maintenance costs will be “outsourced” to the end-users, potentially leading to an even more affordable access to high data rates. An interesting question is therefore which type of services can be offered in these user-deployed networks, and what are the infrastructure densities needed for achieving an “acceptable” user perception. An agent-based “smart” delivery of streaming multimedia and VoIP services was proposed, for cellular networks, in [7], with the purpose of exploiting context information to maximize both user satisfaction and network utilization. In this paper, focusing only on non-interactive media delivery, we propose a specific agent-based pre-fetching strategy, designed for hiding the sparsity of a “spotty” coverage network to the end users, thus, improving their service perception. Some initial work has been presented in [4], analyzing the provision of services with different tolerance to delay in a single user scenario. In this paper, we instead focus on a multi-user scenario, where both the effects of competition for resources at the APs and terminal cooperation are considered. Terminal cooperation as a way to improve delay in “spotty” coverage networks has been initially presented in [6], where mechanisms for sharing “popular” content among users of an infostation system have been investigated. Focusing on characterizing analytical bounds on the content delivery delay for delay tolerant applications, simplified radio channel and service models have been adopted. In this work, instead, we introduce a novel service model that allows us to span different dimensions of the “service space”, bringing time critical aspects into the the investigation. II.
I NTRODUCTION
A fundamental prerequisite, for the success of future wireless broadband systems, is the control of infrastructure and maintenance costs. Due to the fact that the cost of a wireless system with full coverage rises linearly with the provided bandwidth [1], an affordable way to go seems to provide only partial coverage to high data rates. In recent literature, several paradigms considering “spotty” coverage networks have been proposed. The infostation concept (e.g. [2]) is an example of an operator-deployed network providing isolated “pockets” of high bandwidth connectivity. But, in recent years, the wide spread of Wireless Local Area Network (WLAN) APs among private users, is posing a new challenge: being able of opening up to public access the collection of these high-speed APs would create a user-deployed network [3] that might be capable of supporting interesting services [4]. Unlikely to infostation, this paradigm will be exposed to uncoordinated deployment (many different “local” operators) and therefore to a less efficient utilization of resources. On the other hand, in these c 1-4244-0330-8/06/$20.002006 IEEE
P ROBLEM S TATEMENT
The main goal of our work is to investigate whether noninteractive services, characterized by different degrees of “time criticality” can be successfully provided in user-deployed networks. In particular, we want to characterize, for some typical media delivery services, how dense these infrastructures need to be, in order to deliver acceptable user perception. Finally, the quantification of the impact of peer-to-peer user cooperation, on both infrastructure requirements and user satisfaction, is addressed in the case of “popular” content delivery. III. A.
M ODELS
Service
The provision, to nomadic users, of non-interactive “infotainment” content, characterized by different degrees of timecriticality, is investigated in this work. In particular, we focus on the delivery of an infotainment program, here defined as the “collection” of a number of audio/video files called segments.
The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)
Furthermore, according to its time scale validity/variation, each program segment can only belong to one of the following two classes: pre-recorded or live. A segment that has been created before the program download availability, and that remains “unchanged” while the program is played, is defined as “pre-recorded”. According to this definition, “podcasting” (e.g. see [8]) is a type of infotainment program constituted by a single pre-recorded file. On the other hand, some segments might not be available before the beginning of the program, and some could be modified during the program; in our work, these are defined as “live” program segments. The currently available Internetbroadcasted radio programs are, according to our model, infotainment programs constituted by only live segments. On the contrary, if the radio program content remained available for downloading, even after the end of the live show, then it would become a pre-recorded file. In order to classify different infotainment programs, and evaluate their provision in “spotty” coverage networks, we propose a model, where a service S is univocally determined as the following vector: S = {NL , NR , FL , FR , TL , TR , Rc , TP , P }, where NL and NR are, respectively, the number of live and pre-recorded segments in the program, and FL and FR are, respectively, the average size in bits of the live and pre-recorded segments, while TL and TR represent their average playtime duration in seconds. At the same time, TP = (NL TL +NR TR ) is the total playtime duration of the program, while Rc corresponds to the percentage of the program composed by live segments. Thus, if Rc = 0 the program is only composed by one or many pre-recorded files, while if Rc = 1 all files considered are live. An example of service with Rc = 0.2 is shown in Figure 1b. Finally, the parameter P is introduced to model the probability that different users subscribed for the same program (content popularity). This parameters “tunes” the impact of the peer-to-peer information dissemination, meaning that with P = 1 all users are interested in the same service, while with P = 0, all users have completely different interests. In this work, three types of infotainment programs are considered: music clips, news updates and podcasting radio. A “music clip” program (SM ), is assumed to be entirely constituted by pre-recorded video-clip segments (e.g. an mp3 playlist). Under our assumptions, the order in which the songs are played is not important to the end user. The “news updates” (SN ) is instead an audio/video program providing a “breaking news” type of coverage. A typical segment structure is illustrated in Figure 1c. In this work we assume a moderate program duration, and that every TL seconds a new live segment (containing the latest TL seconds of the show) is released. In this case, the order in which segments have to be played is univocally determined. Finally, the “podcasting radio” program (SP ), is modelled as a collection of music files periodically alternated to news update segments. The music files are assumed to be pre-recorded, and “periodically” alternated to segments constituting the “live radio show”. An example is illustrated in Figure 1b.
3 min
(a)
(b)
1
2
(c)
1
2
3
4
5
6
Figure 1: a) Segment structure of SM , a typical music clip program. b) Segment structure of SP , a typical podcasting radio program. c) Segment structure of SN , a program providing news update coverage. B.
Access to Information
The currently evolving network technology is likely, in the near future, to lead to a disintegration of today’s vertically integrated service markets, allowing to select different operators for providing different networking functionalities [9]. Therefore, in this work, we assume that radio access and content provision are two separate markets. This means that users have a subscription with a content-provider, whose remote server location is therefore known to their terminal agents. At the same time, for accessing this content, terminal agents equipped with multiple interfaces, might be able to select different networks, and/or different operators, according to some pre-set user preferences. In this work, we assume that, due to cost-saving motivations [3], only user deployed APs can be chosen for performing radio access operations. Furthermore, a segment can only be accessed in one of the two following modes: networked content access or peer-to-peer content exchange. A networked access to a given segment means that terminal agents, entering within communication distance with an user deployed AP, can reach the wanted file on the content server. On the other hand, if the program is popular, there is a non zero probability that whenever two terminals are in radio contact, one of their two agents has already received, and locally stored, a given segment required by the other. Thus, program segments could be exchanged, in a peer-to-peer fashion, between any two terminals. Obviously, the efficacy of both content access methods depends, for a given mobility model, on both the AP and user densities (respectively ρAP and ρU ). In fact, for the direct access case, the larger the number of APs, the more frequent opportunity of accessing the remote content server location. At the same time, with large user density, there is higher competition for resources at the APs, thus, reducing the number of segments that can be pre-fetched when in coverage. On the other hand, for given content popularity and mobility model, the impact of peer-to-peer segment exchange increases with user density, and diminishes at high AP densities, when the primary access to information is the networked one. The adoption of a pricing strategy discriminating between networked and peer-to-peer content accesses (e.g. similar to [10]) might have a impact on performances, by requiring different content access strategies; however, these additional pricing aspects are left out from the scope of our investigation.
The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)
C.
User Service Perception
The aforementioned programs are assumed to have a publishing time known, in advance, to both users and terminal agents. Once the information is made available on the content server, its delivery to the end user is modelled as a two-phase process coordinated by the software agent. The first phase consists of the pre-fetching of a given number of program segments (eventually all), while the second one is the actual information consumption, or the audio/video program activation on the user device. Moreover, these two phases can be partially overlapped, meaning that consumption can be initiated even before that all the program’s segments have been pre-fetched. Assuming that the pre-fetching phase starts with the program release, the task of the agent is then to decide on when notifying its user that the program is ready to be played. This is done with the overall goal of maximizing the user service perception, which, in this work, depends on two “contrasting” factors: the starting time offset (Tof f ) and program interruptions. In fact, on one hand, the sooner a program is played (low Tof f ), the larger is the user appreciation of it. On the other hand, an earlier start can be achieved by pre-fetching only a limited portion of all the segments constituting the program. This, in turn, might lead to experience some “annoying” program interruption, if the terminal agents cannot access, the remaining program segments before their scheduled playtime. A program is terminated, when all its segments have been received and played. The sum of the time duration of all experienced interruptions is indicated as Tout , while the minimum value of Tof f that allows avoiding ∗ any program interruption (Tout = 0) is indicated as Tof f. Under the assumption that program interruptions have a more negative impact on user perception than a larger starting time offset, we can define the perception of service S experienced by a nomadic user, in a given network configuration (ρAP and ρU ) with the following random variable: Qu (S, ρU , ρAP ) =
∗ Tof f.
∗ Note that for values of Tof f ≤ Tof f , increasing Tof f leads to a decrease of Tout of the same amount, keeping their sum ∗ constant and equal to Tof f . An example of this is illustrated in Figure 2, for the program shown in Figure 1a. This means that Qu can be used to characterize the perception of a given service, under fixed system settings, without the need for con∗ sidering the individual user preferences on how Tof f should be allocated between Tout and Tof f . Furthermore, due to the fact ∗ that Tof f corresponds to the case where no interruptions occur, the evaluation of the packet arrival jitter and the design of strategies for hiding, or reducing, the impact of interruptions on user perception (e.g. playing another locally stored segment) can be left out of this investigation.
IV. A.
N UMERICAL E XAMPLE
Performance Measures
The performance measures, adopted in this work, take into account both the perspectives of the individual users and of the network resulting from the collection of user-deployed APs.
Number of received songs
(a)
t Buffer Status
(b) Tout
T1off
t Empty Buffer
Buffer Status T*off
(c) Qu
t
Figure 2: a) Number of received songs as function of time. b) Buffer
1 ∗ status as function of time. Note that Tof f = Tof f < Tof f brings ∗ program interruptions. c) Buffer status when Tof f = Tof f : no inter∗ 1 ruptions occur. Note that Qu = Tof f = Tof f + Tout is constant.
Assuming a user perspective, for given service S and a system setting, we are interested in the distribution of Qu among the different users. In particular, we targeted the 90-th percentile of Qu , as an indicator of the worst-case performance. This is referred as “time cost” (τ ), and defined in the following way: τ (S, ρU , ρAP ) | [ Pr {Qu (S, ρU , ρAP ) ≤ τ } = 0.9 ] . In particular, it is interesting to quantify how the network size influences the time cost, when only networked content access is considered. This, means to evaluate Qu , as function of ρAP , for a fixed ρˆU and for a service Sˆ with P = 0. The impact of peer-to-peer segment exchange on service perception is quantified by considering two different measures. On one hand, we introduce the “cooperative time cost reduction” ∆τ , representing the ratio between the time cost obtained for a fixed ρAP = ρˆAP when a given service Sˆ has some degree of popularity (P = Pˆ > 0), and the case in which the same service has P = 0. This is a function of ρU and it is represented by the following equation: ˆ ˆ ˆ ρˆAP , Pˆ ) = τ (S, ρU , ρˆAP |P = P ) ∆τ (ρU , S, ˆ τ (S, ρU , ρˆAP |P = 0) On the other hand, the peer-to-peer segment exchange might reduce the AP density needed for providing a given τ . To capture this aspect, we introduced the “cooperative infrastrucˆ Pˆ , τˆ), which is function of ρU ture reduction” ∆ρAP (ρU , S, and it is defined as the ratio between the infrastructure density needed for achieving a fixed τˆ, for a given nomadic user density ρU , when a given service Sˆ has (P = Pˆ > 0), and the case in which the same service has P = 0.
The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)
SM SN SP
NR 10 0 8
FL = FR 11.5 M b 11.5 M b 11.5 M b
TL = TR 180 s 180 s 180 s
Rc 0 1 0.2
P 0|1 0|1 0|1
Table 2: τ and ρeq AP
τ
τ (SM ) | ρAP = 100 τ (SN ) | ρAP = 100 τ (SP ) | ρAP = 100 ρeq AP
Simulation Settings
In order to model lack of coordination, in the deployment of the user-deployed network, the AP locations are obtained as realization of an uniform distribution of parameter ρAP over a service area equal to 1 Km2 . Nomadic users, generated with distribution of parameter ρU , move in the service area according to a mobility model identical to the one described in [4]. Whenever a user enters in the communication range of an AP, its terminal agent connects to the remote server and starts the download of the missing segments. The segments are downloaded according to their progressive identification number (see Figure 1), and newly released live segments have a delivery priority on the pre-recorded ones. The propagation and transmission characteristics of all simulated APs follows the IEEE 802.11g model presented in [4]. At the same time, “backbone limitations”, as well as delays introduced by both authentication and connection setup phases, are neglected. When two or more terminals are at the same time within an AP’s communication range, all its radio resources are instantaneously assigned to the terminal with highest C/I. Furthermore, segments that have been only partially downloaded are discarded when exiting from an AP’s coverage area. Peer-to-peer segment exchange is possible whenever two terminals are located within rpeer meters. The selection of rpeer is subjected to a tradeoff: if rpeer increases, on one hand, the average number of “encounters” between two or more terminals increases, while on the other hand, due to the data rate dependency on distance, the average amount of bit exchanged decreases. In this work, we selected rpeer = 15 meters, which experimentally provides about 95 % probability of exchanging at least one complete segment, and, at the same time, allows between 0.25 and 0.5 peering encounters per terminal every minute, when ρU varies between 20 and 100 users per Km2 . For a given investigated service S, ρAP and ρU , we computed Qu (S, ρAP , ρU ) considering different realization of the APs’ locations and initial user positions, under the assumption that at time t = 0, the program started to be available on the remote server. In particular, in order to represent an “early adoption” scenario, in which both few users have a service subscription and few APs are “shared”, the investigated ρAP ranged between 1 and 100 APs per Km2 , while the investigated user density varied between 1 and 160 users per Km2 . V.
R ESULTS
Three different programs have been selected, in this work, for assessing the feasibility of their provision in user-deployed networks. The segment structure of each of them is shown Figure 1, while their parameters are specified in Table I.
ρAP | τ (SM ) = 310 ρAP | τ (SN ) = 560 ρAP | τ (SP ) = 350
P =0 ρU = ρ¯U 310 540 360
P =1 ρU = 20 291 507 338
P =1 ρU = 160 229 334 252
P =0 ρU = ρ¯U 100 100 100
P =1 ρU = 20 93 93 87
P =1 ρU = 160 64 45 47
650
600
550
500 τ [s]
B.
NL 0 6 2
Table 1: Service Parameters
450
400
350
300 50
SM SN SP 60
70 80 ρ AP [AP s/km 2 ]
90
100
Figure 3: Time cost as function of the AP density for the three investigated services when P = 0. First, we start by illustrating how similar interests between users can concur to improve user service perception (i.e. lower τ ), or to reduce the infrastructure density required for achieving a fixed τ value. In the upper part of Table II, we show the τ s achieved by the three programs in different scenarios. In particular, we compare the “extreme” content popularity cases. In the P = 0 column, the illustrated τ values are the averages of the τ s experienced with all investigated ρU s (this is indicated by ρU = ρ¯U ). Instead, when P = 1, we considered both the cases ρU = 20 and ρU = 160 users per Km2 . For all programs, the time cost reduction is significant. In particular, services with high Rc seems to benefit the most from user cooperation. Instead, in the lower part of Table II, we reported, for all the aforementioned services, the equivalent AP density (ρeq AP ) needed for achieving the τ s of the corresponding non cooperative cases (P = 0). The gains in ρeq AP are large at high user density, meaning that a significant AP reduction can be achieved: approximately 50% less APs are needed, when ρU = 160, for the most “time critical” programs, e.g. SN . The “absolute” time costs, for all the three services, when P = 0 are illustrated in Figure 3, as function of ρAP . All programs achieve improved user perception when more userdeployed APs join the network. In particular, in all points SM has the lowest τ s, while the services with live segments show the greatest percentage of improvement. At the same time, the
The 17th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’06)
100 SM SN SP
95 90
∆ τ [%]
85 80 75 70 65 60 0
20
40
60 80 100 ρ U [users/km 2 ]
120
140
160
Figure 4: Cooperative time cost reduction for the three services when ρAP = 100 APs per Km2 . The gains are relative to the three τ s illustrated in Figure 3 for ρAP = 100 APs per Km2 .
given number of terminal encounters, no additional exchanges are performed. On the other hand, at high Rc s, new segments are frequently created, leading to a more effective and frequent utilization of peer-to-peer segment exchange. Figure 5 shows, for a “target” time cost of 9 minutes, the degree of cooperative infrastructure reduction achievable with different densities of users with coinciding interests (P = 1). It worth noticing that for all services, ∆AP decreases linearly with increasing ρU , and that services SM and SP performs almost identically and both better than SN . This means that having some additional cooperating users can compensate for a reduced infrastructure density and that this equivalence is constant for each investigated service. Taking into account that different ρAP s are needed, by the different services, for satisfying τˆ = 540 s when P = 0 (see Figure 3), we found that about 4.5, 3 and 4 cooperating users are “equivalent” (in the investigated scenario) to a single AP, respectively for services SM , SN , and SP . VI.
100 SM SN SP
90
∆ AP [%]
80
70
60
50
40
30 0
20
40
60 80 100 ρ U [users/km 2 ]
120
140
160
Figure 5: Cooperative infrastructure reduction for the three services, when τˆ = 540 seconds. The reference ρAP s are 60, 100 and 65 APs per Km2 , for respectively SM , SN and SP .
C ONCLUSION
A novel framework for modelling non-interactive “infotainment” services with different degrees of “time criticality”, has been presented and utilized for investigating the user service perception in “spotty” coverage networks. The results showed that already with moderate AP densities, user-deployed networks can deliver acceptable service perception, especially for services with low time criticality. Furthermore, whenever a “critical” mass of users has common interests, the adoption of a peer-to-peer exchange strategy brings the significant gains of infrastructure reduction and/or improved service perception. This suggests that peer-to-peer information exchange is probably the most effective mean for accessing media content in very spotty coverage networks. Thus, even if users do not share the same interests, the confederation of these user-deployed APs should incentivate (even economically) terminal agents to prefetch not only those information segments “interesting” to their own users, but also some additional ones that are likely to be useful to other terminal agents located in the same service area. R EFERENCES
largest τ values are achieved by SN , although it has fewer segments than the other services. This shows that the amount of live behavior (Rc ) has the largest impact on service perception. The cooperative cost and infrastructure reduction gains are illustrated, respectively, in Figure 4 and Figure 5. In particular, the results concerning ∆τ (Figure 4) are achieved with P = 1 and ρˆAP = 100 APs per Km2 , while the ones concerning ∆AP (Figure 5) are specific for the case P = 1 and τˆ = 540 s. However, very similar trends have been found for a large set of different ρˆAP and τˆ reference values. In Figure 4, it is shown that whenever ρU ≤ 60 users per Km2 all programs have similar time cost reduction, while for higher ρU values the peer-to-peer segment exchange starts to be more effective for high Rc services. At the same time, the cost improvement of services with low Rc begin to saturate. This behavior depends on the fact that, for low Rc s, smaller amount of information is periodically created, and therefore beyond a
[1] J. Zander. “On the cost structure of future wideband wireless access”, In Proc. IEEE VTC, 1997. [2] D.J. Goodman et al., “Infostations: a new system model for data and messaging services”, In Proc. IEEE VTC, volume 2, 1997. [3] J.M. Pereira, “Fourth generation: Now, it is personal!”, In Proc. IEEE PIMRC, 2000. [4] J. Hultell, P. Lungaro and J. Zander, “Service Provisioning with Ad-Hoc Deployed High-Speed Access Points in Urban Environments”, In Proc. IEEE PIMRC, 2005. [5] E.C. Efstathiou, G.C. Polyzos, “A Peer-to-peer Approach to Wireless LAN Roaming” in Proc. of WMAS, 2003. [6] W.H. Yuen et al, “Exploiting data diversity and multi-user diversity in mobile infostation networks”, In Proc. IEEE INFOCOM, 2003. [7] T. Kanter, et al., “Smart Delivery of Multimedia Content for Wireless Applications”, in Proc. of MWCN, 2000. [8] www.cnn.com/services/podcasting/. [9] T. Giles, et al., “Cost Drivers and Deployment Scenarios for Future Broadband Wireless Networks- Key research problems and direction for research”, in Proc. of IEEE VTC, 2004. [10] M. Lindstr¨om, P. Lungaro, “Resource delegation and rewards to stimulate forwarding in multihop cellular networks”, in Proc. of IEEE VTC, 2005.
