Sie sind auf Seite 1von 7

A Social Group-aware and Tunneled Scatternet Formation Scheme

Myeong-a Yang, Yangwoo Ko, Dongman Lee


Information and Communications University
{may, newcat, dlee}@icu.ac.kr

Abstract performance will be improved. Manish et al. [8]


introduce a Communicating Group (CG), defined as a
Bluetooth is a major technology for short range group of mobile devices with frequent data transfers
wireless communication. Scatternets expand its use to amongst themselves. CGs are identified by observing
larger networks. If communication patterns are known traffic and finding frequently communicating peers.
before scatternet formation, frequently However, this scheme does not work well in a
communicating pairs could be connected with fewer pervasive computing environment where CGs can
hops to enhance performance. Communication emerge and disappear dynamically due to spontaneity
patterns are obtainable by observing traffic, but it is of interactions because it is difficult to reliably
ineffective in a pervasive computing environment estimate the communication patterns. In a pervasive
where interactions are mostly spontaneous. We computing environment, interactions between two
propose to use social group membership as an specific people may last for only a short time.
estimate of communication patterns. The proposed interactions between people happen spontaneously,
scatternet formation scheme forms a scatternet per a and may last for only a short amount of time especially
social group and then connects scatternets through between infrequently communicating peers. In this
tunnels. In order to ensure that each communication case, identifying a CG for them may take similar to or
pair has a reasonable share of bandwidth, the longer than the length of their communication session
proposed scheme considers the numbers and the and then the identified CG becomes useless. .
positions of the tunnels. Simulation results show that In a real world, a person in general belongs to a
the proposed scheme not only enhances the total social group and interacts with others who belong to
throughput of scatternets but also shares network the same social group more often than those who do
capacity more fairly among communication pairs. not [2]. In this paper, we propose an efficient
scatternet formation scheme exploiting social group
1. Introduction membership as an approximation of a CG. Even
though interactions between two specific people will
Bluetooth [1] enables cell phones, PDAs, and not last for long time, communication with the same
notebooks to be connected without wire and is used to social group member may take place with higher
form a Wireless Personal Area Network (WPAN). The possibility than that with other group members. We
basic communication unit of Bluetooth is a piconet assume that each person knows which social groups he
that consists of one master and up to 7 slaves. To or she belongs to. Devices belonging to the same
connect more than 8 Bluetooth devices, multiple social group form a piconet or a scatternet using the
piconets are formed as a scatternet and several most well known scatternet formation scheme, TSF
scatternet formation schemes have been proposed [6]. When there is a need to interconnect a social
[3][4][5][6][7][8]. One of the important goals in group with other social groups which usually lasts for
scatternet formation is to maximize the performance a shorter amount of time than interactions between
of a scatternet such as the capacity of a scatternet intra-group members, a set of tunnels is selected using
[4][7]. the topology information of the constituent scatternets
Most of scatternet formation schemes form a and interconnects them through the tunnels unlike the
scatternet by connecting all devices within an area, existing schemes in which the scatternets are merged
which may cause frequently communicating pairs to into a single scatternet. When selecting tunnels, the
have longer path lengths than optimal ones. If devices proposed scheme takes into consideration the number
are clustered to form a collection of smaller sized of hops and branches and traffic distribution. Our
scatternets or piconets, frequently communicating scheme not only enhances the total throughput of
pairs will have shorter average paths so the scatternets but also shares network capacity more
fairly among communication pairs. The simulation
results show that our scheme achieves higher average themselves, and which forms different piconets for
TCP throughput and shows a smaller variance than different CGs. This allows simultaneous
TSF. communication in different CGs and hence leads to
The remaining sections are organized as follows. higher throughput, lower delays and less packet drops.
Section 2 briefly explains related work. Section 3 This scheme analyzes traffic flow patterns to identify
introduces a sample scenario and describes what we CGs. It takes relatively long time compared to
should consider when forming a social group aware communication session in pervasive computing
and tunnel based scatternet. The proposed scheme is environment so identified CGs becomes obsolete
described in section 4, and we show the performance quickly. Moreover, even though [8] suggests the use of
of our scheme’s performance in section 5. Finally, CGs, it does not propose a concrete scheme for
section 6 concludes the paper. forming and adapting piconets to the current state of
CGs.
2. Related Work
3. Design Consideration
2.1. Social Group
3.1. Scenario
A social group is “a number of individuals, defined
by formal or informal criteria of membership, who Our scenario is extended from that of [2]. Several
share a feeling of unity or are bound together in students are taking a course and they are divided into
relatively stable patterns of interaction” [2]. A person three groups for a term project. Let’s call them group
can become a member of social group when there is A, B, and C. We assume that all students have
consensus among group members. Bluetooth enabled devices. During the class, the
The main concerns in [2] are the relation between a professor asks the students to discuss a subject related
group member and his or her devices, the format of to the class with group members and to present the
user group profile, and membership management. It relation of their term project with this topic. Groups A,
does not consider how to efficiently facilitate B, and C form individual networks and start
communication among social groups, which is the discussion. While discussing, group members
main theme of our work. exchange related data and participate in a
collaborative review process while preparing the
2.2. Bluetooth and Scatternet Formation presentation material. After discussion, the three
groups need to be interconnected into a large network
The Bluetooth Specifications 1.2 [1] defines the in order to enable students to exchange presentation
operation of Bluetooth and its protocol. Bluetooth materials or give comments in the middle of a
differs from a contention–based protocol such as IEEE presentation. In addition to the inter-group
802.11, since Bluetooth devices form master-slave communication, there is also continuous intra-group
links and each slave adjusts its communication communication for sharing thoughts with the others in
frequency hopping sequence to that of its master. the same group.
Bluetooth adopts a time division duplex
communication scheme, where the master decides 3.2. Social Group-aware Scatternet
which slave communicates with itself in next time
slot. A master has up to 7 slaves and they form a Manish et al. [8] propose to observe traffic among
piconet. A scatternet is proposed to interconnect more nodes in order to group a collection of nodes in a
than 8 devices, but it has not been standardized yet. space into subgroups of frequently communicating
There have been several scatternet formation nodes. However, interconnections between identified
schemes such as Bluemesh[3], Bluenet[4], Shaper[5], subgroups may last for only a short time due to
and TSF[6]. Scatternet formation schemes can be spontaneity of interactions and CGs for them quickly
classified by the resulting topology (for example, tree become obsolete in a pervasive computing
or mesh), by the type of bridge node (for example, environment. We need an alternative way to find
master-slave or slave-slave), and by whether all nodes communication patterns that are stable over time.more
are in transmission range or not. Most of the existing efficiently. In a real world, as exemplified above, a
scatternet formation schemes focus on the formation of person usually belongs to a social group and
one scatternet without considering social relationships communication mainly happens among social group
among devices (or their owners). members for a considerable amount of time for
Manish et al. [8] proposes the use of a collaborative activity [2]. Therefore, Instead instead of
Communicating Group (CG), defined as a group of forming a single scatternet, it is more appropriate to
mobile devices with frequent data transfers amongst form a separate scatternet for devices belonging to the
same social group and provide a way to connect
corresponding scatternets when there is a need of
interactions across social groups. 240
To support communication across social groups,
220
their corresponding scatternets need to be
interconnected. We call this interconnection as a 200
tunnel. Communication pairs across social groups 180
have longer paths than that of within groups, so they 160

Throughput (Mbps)
will show lower throughput. The path length of an
inter group communication pair can be reduced by 140

adding more tunnels. For example, if there are two 120


string (linear?) topology whose numbers of nodes are 100
100 (too big!); two tunnels that composed of both 33rd
80
nodes and both 66th nodes shorten average path
lengths than one tunnel that consists of both 50th 60
nodes. However, too many tunnels decrease the 40
capacity of the entire network because an additional 20
tunnel increases the number of branches per node. So
0
the number of tunnels should be such that the path
Communication pairs
lengths of inter-group communication pairs are short
(short 라고 하면 너무 모호하지 않나? optimal 이라
고 하면 안되나? 경로 길이를 가장 짧게 해주는 것 Figure 1. Distribution of TCP throughput per
은 아니니, optimal 이라고 부르기는 힘들다고 생각 communication pairs
합니다 bandwidth 관점에서 보면 optimal 한 거 아
닌가? 즉 tunnel 이 너무 많아서 bandwidth 가 줄어 4. Proposed Scheme
들지 않는 범위 안에서 최대한 패스를 짧게 하는
것이라면 그런점에서는 optimal 한 것이죠. 우리 This section describes an overall structure of the
스킴이 그런지 아닌지는 모르겠고) while the proposed scheme and the metric that is helpful for
network capacity is not reduced significantly. choosing tunnel nodes. We will then explain the
We must consider the position of each tunnel as proposed scatternet formation scheme in detail.
well because it determines traffic distribution. Tunnels
forward inter group traffic from one scatternet to 4.1. Assumptions and overall structure
another, so all inter-group traffic is concentrated in
tunnel nodes and their neighbors. If they also forward We assume that each person knows which social
a majority of intra group traffic, they may become a groups he or she belongs to. Devices that belong to the
bottleneck. So the tunnel nodes should be selected same social group form a piconet or a scatternet using
from the nodes in which intra-gorup traffic is less an existing scatternet formation scheme. Our current
concentrated. implementation uses TSF [6]. As exemplified in the
Figure 1 shows the distribution of TCP throughput scenario given in Section 3.1, there is a need to
per communication pair in a typical scatternets of 20 interconnect a social group with other social groups.
nodes. TCP throughputs are sorted in a decreasing We also assume that this information of when and to
order. As you see, a few communication pairs use the which social group they should be connected is known
majority of network capacity, while most pairs utilize to member nodes out of band. A Representative of
very low bandwidth. Even though this problem is each scatternet establishes a link to share information
inevitable except for star topology, we need to alleviate needed for forming tunnels. The election of a
this unfairness. When determining the number and the representative node may be different depending on a
positions of the tunnel, we should consider fairness scatternet formation scheme. For easier
among intra-group traffic as well. implementation, we assume that representative nodes
are predefined. This assumption is quite reasonable
because a social group usually has a leader (for
example, the term project leader in our scenario).
After establishment of a link, we can select a set of
tunnels based on the shared topology of the constituent
scatternets. As described Section 3.2, we need to
consider the number and the positions of the tunnels.
For this purpose, we define a metric that estimates the
performance of a scatternet, which we will describe in
Section 4.2. Each representative node evaluates the C D
metric and finds the best tunnels as described in x y z
Section 4.3.

4.2. Proposed Metric (b)


Figure 2. Link capacity per CP
Our metric calculates the lower bound of the
maximum throughput that a communication pair can Based on weighted link capacity per CP, we can
have. A scatternet that maximizes this metric also obtain the total network flow. It is a sum of the
maximizes the throughput of traffic that has the lowest individual flow of each source-destination pair. An
throughput, for example, inter-group traffic that passes individual flow means how much capacity a pair can
a bottleneck. Therefore, a scatternet with a high metric occupy for a given capacity. The individual flow of a
value minimizes variations of the throughput of pair is the minimum link capacity per CP on their
communication pairs and shares capacity more fairly. routing path. Since a longer path increases an error
The proposed metric considers the following: 1) the rate, throughput decreases as a path length increases.
path lengths of communication pairs (or CP So the total network flow is divided by an average hop
hereinafter), 2) the number of branches per node, and count.
3) the link capacity per CP. Since a node utilizes one In summary, the proposed metric can be written as
link at time, overall performance degrades if we follows.
increase the number of branches to reduce the average TotalNetworkFlow
path length. Meanwhile, if we reduce the number of Metric =
AverageHopCount
branches per node, latency becomes longer due to the
increase of the average path length so that overall ∑ PathLengthOfCP i

throughput also degrades. Link capacity per CP is AverageHopCount = i

NumberOfCP
defined as a reciprocal of the number of CPs whose
traffic passes through the link. A link that has the TotalNetworkFlow = ∑ IndividualFlowOfCPi
i
lowest link capacity per CP becomes a bottleneck in a
IndividualFlowOfCPi = min Link j CapacityPerCPi
scatternet because the performance of each CP will be j
bound by a link with the lowest link capacity.
Calculating link capacity per CP is done in two
4.3. Tunnel Selection Scheme
stages. In the first stage, we count the number of CPs
whose traffic passes a link and take an inverse of it.
In this section, we will explain a scatternet
For example, in Figure 2(a), if we assume that all
interconnection procedure that selects tunnels between
source-destination pairs send data, there will be three
two social groups wanting
CPs: (u, v), (u, w), (v, w). (u, v) and (u, w) will use
interconnection.(interconnection procedure 에 대한
link A while (u, w), (v, w) will use link B. Therefore
설명을 넣으라는 것이 아니라 그 procedure 가 어
the link capacity per CP of A is 1/2 and that of B is
떤 형태로 구현되느냐 하는 것을 교수님이 질문하
also 1/2. In the next stage, we calculate the weighted 신 것 같습니다. 즉 그 프로시져가 어떤 응용 프로
link capacity per CP that considers the number of 그램의 일부인지 아니면 프로토콜 스택의 일부인
branches. The weighted link capacity per CP is the 지. 만약 후자라면 누가 그것을 trigger 하는지 등
link capacity per CP divided by the number of 을 설명해야 한다는 것이죠. 아래에도 그냥 그림
branches of neighbor nodes. Between two neighbor 만 들어 있는데 추가 또는 변경되는 구성요소가
nodes constituting a link, we only consider the node 뭔지 그리고 그 역할이 무엇인지 얘기하지 않으면
with more branches. (Here, we assume that each node 이해하기가 힘들겠지요.) If two social groups need
spends the same amount of time with any link.) In to be connected, one of the users such as the leader of
Figure 2(b), since node x and node y has two links, if a group initiates the interconnection procedure with
the load per link for link C is 1/8, its weighted load his or her Bluetooth device. This device becomes a
per link is (1/8)/2 = 1/16. Since node y has 2 links representative node of a group. The interconnection
while node z has 3 links, if link D’s load per link is procedure sends a connection request to the scatternet
1/9, its weighted load per link is (1/9)/3 = 1/27. formation component of the Bluetooth stack with the
A B counterpart scatternet’s identifier. The same procedure
u v w happens in the counterpart scatternet. The scatternet
formation component starts an inquiry procedure with
(a)
Bluetooth which results in a link between the two
representative nodes. One of these nodes that receives
an inquiry packet becomes a master representative
node and the other one does a slave representative representative nodes should be removed except when
node. it is a tunnel.

5. Evaluations
Yangwoo Ko @ ICU 님의 말:
Maneshi 논문에서 CG 부분은 달랑 한두 패러그래
프 뿐이고
Yangwoo Ko @ ICU 님의 말:
아이디어 차원에서만 얘기한거라 구체적으로비교
하기가 사실상 어렵다는 얘기를 좀 넣으면 어떨지
요?
HCI (and scatternet formation
procedure) thodlee 님의 말:
고양우씨가 제시한 것은 perf evaluation 에서 얘
기하면
thodlee 님의 말:
되겠네요

5.1. Performance comparison of group-aware


Figure 3. Bluetooth protocol stack scatternets and group-oblivious scatternets

The scatternet formation component is notified via Our simulation is based on Blueware [9] which
a callback at the end of the link establishment implements most of the functions that a Bluetooth
procedure. The callback function of the slave device should provide, and also supports an
representative node sends its scatternet topology to the implementation of TSF. Table 1 compares two cases:
master representative node. Discovering the scatternet group-aware scatternet and group-oblivious
topology is out of this paper’s scope, but this scatternets. A group-aware scatternet refers to a
information is easily obtainable. For example, all scatternet that is an interconnection of a set of socially
nodes know the overall topology in SHAPER [5], and grouped devices while a group-oblivious scatternet
the root remembers the topology in TSF. refers to a scatternet that is formed by connecting all
Upon receiving the topology of the counterpart devices within an area. The group-aware scatternet
scatternet, the master representative node computes formation simulations start with 20 Bluetooth devices
which set of tunnels maximizes the proposed metric. and form two small sized scatternets, each of them
After computation, it sends the result to the slave having 10 devices. Two scatternets are then
representative node. Each representative node sends a interconnected by a set of tunnels that have the highest
tunnel establishment command to the first tunnel node metric values. The group-oblivious scatternet
candidate. Two tunnel node candidates establish a link formation simulations start with 20 devices and form a
or a tunnel using the same link establishment different topology according to a different seed
procedure explained above. A tunnel node candidate number. We repeat the simulations with different
reports its success or failure of the tunnel traffic patterns 20 times for a given topology. We set
establishment to the representative node of its group. the number of intra-group traffic twice of that of inter-
After receiving the result of tunnel establishment, the group traffic since we assume that intra-group
representative node initiates additional tunnel members will communicate more frequently. This
establishment if needed. Otherwise, it ends the traffic pattern is applied to the rest of the paper.
tunneled scatternet formation. Group-aware scatternet records a higher average
As described above, tunnels are established one TCP throughput for intra-group traffic and shows
after another. There are two reasons for sequential lower variance as well. Lower variance means that the
construction of tunnels. First, if several devices throughput distribution among CPs is closer to an
transmit an inquiry packet simultaneously to establish average, so a group-aware scatternet permits CPs to
links, packets will probably collide and thus no link share network capacity more fairly. In addition to
could be constructed. Second, the tunnel establishment overall performance, group-aware scatternets offer
procedure sometime fails due to the absence of a higher bandwidth to intra-group traffic without
counterpart device or for other reasons. Therefore, the sacrificing much inter-group traffic.
representative node should know if tunnel
establishment succeeds or fails so that it can decide Table 1. Performance comparison of group aware
whether to move to the next tunnel setup or to select scatternets and group unaware scatternet
another tunnel candidate. Once all tunnels are
established, a temporary interconnection link between
Group aware Group unaware Communicating Group (CG), a group of devices that
Inter Intra Inter Intra frequently communicate, and a scatternet constructed
group group group group according to CG shows high throughput. Identifying
traffic traffic traffic traffic CGs takes quite amount time compared to the session
Average TCP time in pervasive computing environments and
Throughput 29.1 52.8 34.8 33.7 identified CGs become obsolete quickly. ( 이 에 대 한
(Kbps) 우 리 의 방 법 요 약 ; ) Therefore we propose to use
Standard social group as an alternative of CG. And if there are
deviation for 55.7 65.9 72.1 77.3 occasional needs to communicate across social groups,
all CPs we interconnect them through tunnels. We also
introduce a metric that is useful for selecting a set of
tunnels that interconnect two social groups. Our
5.2. Relation between the metric and TCP
metric takes into consideration the number of hops
throughput and branches, and the link capacity per
communication pair. ( 성 능 평 가 결 과 요 약 ;)
This section shows the relation between the Simulation results demonstrate that a scatternet
proposed metric and the TCP throughput, which is formed by our scheme shows higher performance than
depicted in Figure 3. The y axis is the average TCP group-oblivious scatternets formed by an existing
throughput of inter-group traffic from 20 experiments scatternet formation scheme. We also show that the
and the x axis is the metric value of a given tunneled proposed metric and the TCP throughput have a close
scatternet. We can conclude that the proposed metric and positive correlation.
is suitable for evaluating a tunneled scatternet since We compare the performance between group-aware
the proposed metric and the TCP throughput of inter- scatternets and group-oblivious scatternets, both of
group traffic has a positive correlation. them are based on TSF. Additional experiments are
needed to demonstrate that our scheme is also an
improvement compared to other scatternet formation
schemes besides TSF.

7. References
[1] Bluetooth Specification Version 1.1, Bluetooth Special
Interest Group, http://www.bluetooth.com, February 2001.
[2] B. Wang, J. Bodily, and S. K. S. Gupta, “Supporting
Persistent Social Groups in Ubiquitous Computing
Environments Using Context-Aware Ephemeral Group
Service”, in Proceedings of the Second IEEE Annual
Conference on PERCOM 2004.
[3] C. Petrioli, and S. Basagni, “Degree-constrained
multihop scatternet formation for Bluetooth networks”,
Figure 3. Relation between the proposed metric and IEEE Globecom 2002, Taipei, Taiwan, November 2002.
the TCP throughput of inter-group traffic [4] Z. Wang, R. J. Thomas, and Z. Haas, “Bluenet - a new
scatternet formation scheme”, HICSS-35, Big Island,
Hawaii, January 2002.
6. Conclusion [5] F. Cuomo, G. D. Bacco, and T. Melodia, “Shaper: a self-
healing algorithm producing multi-hop bluetooth
(많은 수의 디바이스가 연결되는 스캐터넷 에 scatternets”, IEEE Globecom 2003, San Francisco USA,
서 의 formation 문 제 점 제 시 ; ) Most of scatternet December 2003.
formation schemes form a scatternet by connecting all [6] G. Tan, A. Miu, J. Guttag, and H. Balakrishnan, “An
devices within an area, so frequently communicating efficient scatternet formation algorithm for dynamic
pairs devices cannot occupy obtain reasonable capacity environments”, in IASTED International Conference on
Communications and Computer Networks, Boston, MA,
due to long average path lengths. (average 가 커진다
November 2002.
고 해서 꼭 나쁜건 아니죠. 자주 통신하는 놈들이 [7] T. Melodia and F. Cuomo, “Ad hoc networking with
가까이 있으면 되니까.  이상적인 CG 의 경우. Bluetooth: key metrics and distributed protocols for
앞의 어디선가 설명한 것 처럼 자주 통신하는 장 scatternet formation”, Ad Hoc Networks, vol. 2, no. 2, pp.
치인데도 서로 멀리 떨어지는 “경우도 있다.” 그 109–202, April 2004.
런 경우에는 문제가 된다. 안그러고 이렇게 얘기 [8] M. Kalia, S. Garg, and R. Shorey, “Scatternet structure
하면 항상 문제가 된다고 과장되고 얘기하는 것으 and inter-piconet communication in the Bluetooth system”,
로 비칩니다.) Manish et al. [8] proposes the use of a
in IEEE National Conference on Communications New [9] G. Tan, “Blueware:Bluetooth Simulator for NS.” MIT
Dehli, India, 2000. Lab. Comput. Sci., Cambridge, MA, October 2002.

Das könnte Ihnen auch gefallen