Adaptive Information Dissemination:
An Extended Wireless Data Broadcasting
Scheme with Loan-Based Feedback Control
Chih-Lin Hu and Ming-Syan Chen, Senior Member, IEEE
Abstract—The dissemination of numerous information broadcast services gives rise to the scalability issue in wireless networks. Previous researchers have shown that the push-based data broadcast mechanism is efficient in reducing message traffic. However, most research efforts are dedicated to the dissemination of static information contents. In practice, information broadcast services can produce and deliver dynamic information contents. To efficiently convey the dynamic data, we devise, in this paper, an adaptive information dissemination mechanism by exploiting the functionality of data broadcasting, to support the dissemination of static and dynamic information services simultaneously. In our design, both static and dynamic information services are subsumed as service groups, i.e., the building blocks with the uniform representation of structure and group popularity and, thus, the conventional scenario becomes a special case of our framework. Furthermore, in order to tolerate the broadcast traffic dynamics, we design an online loan-based slot allocation and feedback control technique to deal with the adaptation of the service group classification, bandwidth allocation, and broadcast schedule so as to avoid performance degradation. It is shown by the experimental study that the proposed adaptive information dissemination mechanism associated with the online loan-based feedback control is able to achieve a substantial reduction of message traffic for dynamic information dissemination in wireless networks.
Index Terms—Push, adaptation, feedback control, data broadcast, information dissemination, wireless network.
æ
1
I
NTRODUCTIONI
Nthe aspects of bandwidth capacity and information flow,the asymmetric communication essentially encounters many challenges in the developments of information broadcast applications and services. Particularly, the capacity of downward bandwidth is larger than the opposite one, for example, in satellite networks, radio networks, wireless links in WLAN, and mobile cellular networks. In addition, information flow can be asymmetric in several examples, such as electronic auction and tender, instant messaging, personalized news distribution, and Web surfing systems, to name a few. Thus, with the rapid proliferation of informa-tion broadcast services and mobile recipients in wireless networks, the traditional pull-based/client-server delivery model suffers from the scalability problem and performance degradation. Substantial researchers have shown that the push-based/broadcast delivery model is a viable solution to resolve the scalability issue, especially in asymmetric communication environments. In the paradigm of push-based data broadcasting, data are pushed in a round-robin manner over a shared broadcast medium and accessed by the clients passively without pull requests [10], [23], [24], [25], [40]. More supplements to the data broadcast model will be reviewed in Section 2.
Notice that, although many previous research works have explored the data broadcast methodologies, most of them are manifested in the traditional data management systems, where a data item, specifically mapped to a pair of state and value in the database, is persistent and static. However, many modern application domains encounter different scenarios where data contents are produced continually and dynamically. In these scenarios, dynamic data streams need to be online processed rather than being stored and later retrieved to answer queries. Indeed, the data broadcast traffic has the nature of dynamic changes. As a result, most prior studies in the data broadcast commu-nity have to resort to either the assumption that the traffic is static, or that the prior knowledge of traffic patterns is available. Hence, these methods are mainly designed for the static optimization of data broadcasting, but not to address the efficient dissemination of dynamic data and information contents.
In this paper, we examine the issue of adaptively disseminating dynamic data contents on broadcast chan-nels. We consider two important phenomena: 1) the data broadcast contents are produced continually and dynami-cally. 2) The data broadcast traffic has the nature of dynamic changes. Accordingly, we have designed a novel group-based information dissemination (GID) mechanism with the loan-based slot allocation and feedback control (LSAFC) technique by exploiting the functionality of data broadcasting, to cope with the impacts of information and traffic dynamics. For brevity, the integration of the GID mechanism and the LSAFC technique is abbreviated as GID+LSAFC. The GID mechanism subsumes static and dynamic information services as service groups, i.e., the
. The authors are with the Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan, Republic of China,
E-mail: [email protected] and [email protected]. Manuscript received 23 Sept. 2002; revised 16 Apr. 2003; accepted 10 July 2003.
For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference IEEECS Log Number 20-092002.
building blocks, representing access commonality. Further, the online LSAFC technique deals with the adaptation of group classification, broadcast contents, and bandwidth allocation to tolerate dynamic traffic changes. Note that the GID+LSAFC mechanism is under the paradigm of a clients-providers-servers system [9], [15]: particularly, items are produced dynamically by the information providers, disseminated efficiently by the broadcasting servers, and accessed in the client sides. In addition, both of the client-oriented and the server-oriented traffic factors are considered. The former includes the request workload, skew access pattern, client population, etc., and the latter indicates the dynamic item production rate, specifically a variable number of dynamic items from an information provider within a cycle time. Note that a dynamic item corresponds to an item produced dynamically, rather than an item whose length varies in time. The GID+LSAFC mechanism analyzes the impact of information and traffic dynamics, while its performance is evaluated online. These features distinguish our work from prior ones, which consider static data (and certain traffic factors) in the traditional data broadcast model [10], [38].
The GID+LSAFC mechanism is described as follows: The GID mechanism aims to mitigate the heavy traffic penalty in response to dynamic information dissemination. The access commonality of an information service is conceptua-lized as a group with the uniform structure. Two group types, an actual group and a virtual group, are used to represent access commonality. An actual group includes a number of clients who retrieve dynamic data items from the same information broadcast service. In contrast, a virtual group is defined to backward support the original data broadcast scenario where clients have an interest in a certain item. In other words, a virtual service always generates the same item accessed by a virtual group as a special case of the GID model. Each group is allocated with a slot quota to deliver dynamic data items. For example, it is desirable to assign some slots to a news broadcast service for distributing updated news at any time. In contrast, the manner of predetermining the broadcast contents and schedules by the classical broadcast scheduling schemes is not suitable in dynamic environments.
When data items are generated dynamically and con-tinually, in order to attain the system robustness, the broadcasting server needs a runtime technique to monitor dynamic traffic changes and to react to them adaptively. Note that the design of the LSAFC technique provides the GID mechanism the ability to adjust the broadcast contents on the broadcast channels. Explicitly, during a broadcast cycle, the LSAFC technique is able to analyze the temporary traffic patterns and manipulate the slot quota dynamically. When a group has not enough slots in response to additional items, a group can loan slots dynamically from other groups with the specific loaning policy. In the end of the broadcast cycle, the broadcasting server will have the feedback from the loan information. Accordingly, the GID mechanism unifies the quantitative measure of message traffic generated by different groups and, therefore, con-trols the adaptation of group classification, bandwidth allocation, and broadcast schedule periodically. Herein, the
above procedure can be better understood by an illustrative example below.
Example 1.The design of the GID+LSAFC mechanism is an
adaptive, hybrid data dissemination framework in dynamic environments. Suppose there are five clients who receive dynamic data from four information
broad-cast services fI1; I2; I3; I4g in a service area. Given
10 broadcast slots, according to the traffic patterns in the last broadcast cycle, the broadcasting server initially
allocates slot quotas fs1; s2; s3; s4g ¼ f1; 2; 7; 0g to the
information services in the beginning of the broadcast
cycle. Thus, the broadcast program contains fI1; I2; I3g.
As S4¼ 0, dynamic data from I4will be delivered in the
pull manner. Suppose that I1is accessed by three, I2 by
two, I3 by two, and I4 by four clients; I1; I2; I3; and I4
dynamically generate/update information with variable
rates f1; 5; 4; 1g within the time interval of the current
cycle. “1” indicates that I
1delivers a single/static item
corresponding to the traditional data broadcast case. In
this context, I1 is regarded as a virtual group, but the
others are actual groups. During the broadcast cycle,
I2loans three slots from I3 to broadcast excess items by
using the loan-based slot allocation. In the end of this cycle, the server recalculates the respective traffic loads of service groups as f3; 10; 8; 4g by multiplying the number of clients and the number of dynamic items in a specific group. Accordingly, the broadcast program in
the next cycle will contain fI2; I3; I4g with slot quotas
fs1; s2; s3; s4g ¼ f0; 5; 4; 1g. Therefore, the GID+LSAFC
mechanism has a total traffic of 14 messages (10 messages
for broadcasting fI1; I2; I3g and four messages for
delivering I4) as compared to 25 messages by using the
exclusive pull mode.
In the data broadcast arena, most research works did not consider disseminating dynamic data contents with respect to information and traffic dynamics. In contrast, the study of this paper involves both client-oriented and server-oriented traffic dynamics in the clients-servers-providers environ-ments. Accordingly, we have designed the GID+LSAFC mechanism, in essence, an adaptive information dissemina-tion framework, which not only preserves the original functionality of data broadcasting, but also offers the dissemination of dynamic information services. Extensive simulations have been conducted to investigate the scal-ability and robustness of the GID+LSAFC mechanism. Also, we have derived new server-centric performance metrics, which provide much insight into the problem of dynamic information dissemination, instead of the client-centric metrics used in the traditional data broadcast model where static data are concerned.
The rest of the paper is organized as follows: Section 2 gives the problem description and related work. Section 3 describes the design of the GID mechanism, and the LSAFC technique is presented in Section 4. Section 5 shows the performance evaluation. This paper concludes with Section 6.
2
P
RELIMINARYThis section first describes the perspective of this paper and then reviews prior works in the data broadcast model.
2.1 Perspective in Dynamic Information
Dissemination
Fig. 1 illustrates a generalized mobile data access frame-work [1]. Due to the lack of an efficient multicast mechanism in the existing mobile data systems, a broad-cast/multicast message incurs a number of message relays and transmission cost within the Network and Switch Subsystem (NSS) and the Base Station System (BSS) simultaneously [1], [26], [27], [33]. With the rapid growth of the information broadcast services and mobile recipients, reducing the enormous communication overhead is a crucial challenge in wireless networks.
In the BSSs, there exist commonly shared broadcast channels. Many prior works show that the use of data broadcasting is scalable in the traditional data management systems [2], [10], [11] where data are persistent and static. However, as mentioned before, in many modern applica-tion domains, data items on the broadcast channels are generated continually and dynamically. Since previous efforts did not explore the dissemination of dynamic data contents, they are not applicable to the dynamic environ-ments. Comparatively, we exploit, in this paper, the potential of data broadcasting for the dissemination of dynamic data contents so as to mitigate the heavy traffic penalty. Therefore, our proposal is able to support not only the dynamic information dissemination, but also the traditional data broadcast scenarios.
2.2 Related Work
2.2.1 Data Broadcast and Periodicity
The basic data broadcast model assumes that an information server maintains a broadcast database and applies a periodic schedule to disseminate data items to clients through a shared medium [2], [23], [40]. In teletext systems, [7] proves that there exists an optimal schedule which is periodic, and also presents the lower bound of mean service
time. Deductively, the greedy heuristics are empirically tested to formulate the suboptimal policy for the determi-nistic optimization problem [37]. However, the data broad-cast problem is NP-Hard [8] as a special case of the generalized maintenance scheduling problem.
2.2.2 Transaction-Based Data Broadcast
The transaction-based data broadcast is first discussed in the Datacycle project [11]. The maintenance of semantic and temporal coherence is the primary issue [32]. The consis-tency in [11] is ensured by the expensive serialization. Shanmugasundaram et al. [35] introduces new correctness criterion. In addition, the invalidation [30] and the multi-version [31] techniques are presented to increase the concurrency of read-only transaction.
2.2.3 Broadcast Disks
The “broadcast disks” introduces a simplified data broad-cast model with restrictive assumptions [2] and, thus, renders the data broadcast problem to be not NP-Hard. Data are static in this model. Each partition of static data set is viewed as a rotating disk, and the relative access popularity determines the rotation speed. The broadcast schedule is the output from these rotating disks. Note that many research efforts of scheduling [29], indexing [14], [22], [25], prefetching, and caching techniques [3], [41] are elaborated upon this context [10].
2.2.4 Broadcast Scheduling Strategy
In the traditional data management systems, whatever the length of an item may be, an item is static and its length is ”constant.” Accordingly, this proposition renders the data broadcast problem tractable. Otherwise, if the length of an item varies at any time, the classical broadcast schedule schemes are not sustainable without guarantees of perfor-mance, reliability, and robustness. Regarding the generation of a push-based broadcast program, Wong’s [40], Su et al.’s [37], and Broadcast Disks [2], [13] consider the static and uniform-length item, the item access frequency (or request rate), and the elapsed span (from the latest broadcast of the item), except Hameed and Vaidya’s [18], [39], where items
have the nonuniform lengths. As for scheduling the pull-based broadcast program, some basic schemes, e.g., Earliest First Request, Most Requests First, and Longest Wait First, are designed to schedule static and uniform-length items, as reviewed in [16], [40]. In [6], the R W algorithm combines the benefits of MRF and EFR with a lower overhead. On the other hand, Shortest Service Time First (SSTF) scheme schedules static items of nonuniform lengths, but favors items of short lengths. Recently, a preemptive scheduling scheme [5], [34] is explored. Particularly, if a lengthy item is dividable, it is profitable to preempt the broadcast of a lengthy item to broadcast another item with a large number of outstanding requests.
2.2.5 Hybrid Data Broadcast Model
In general, an information server can disseminate data in either the pull or the push data delivery model. However, the push delivery mode, which broadcasts all data periodically, can result in an unacceptable access time when the number of data items in the database is huge. In contrast, the pull delivery mode responding to a client’s individual request inevitably incurs a scalability bottleneck with a large number of mobile recipients. Therefore, the hybrid delivery strikes a compromise between the trade offs [4], [12], [19], [21], [36]. Explicitly, data items are classified as hot (or popular) and cold (or unpopular) items according to their access frequencies, and item slots in a broadcast cycle are partitioned into the push and the pull sets. Push slots carry hot items and cold items are delivered on pull slots by the request-response manner. With the access commonality, the hybrid data broadcast model provides an opportunity to make the access time malleable. Conse-quently, in this paper, we deliberate the design of dynamic information dissemination based on the hybrid data broad-cast model.
3
D
ESIGN OFGID M
ECHANISMSection 3.1 describes the system model and notation. Section 3.2 abstracts the design of the GID mechanism. The
parts in the GID mechanism are presented in subsequent sections: group association, group popularity, group classi-fication, broadcast program generation, and group schedule priority.
3.1 System Model and Notation
As illustrated in Fig. 2, the GID mechanism is an extended hybrid data broadcast model for dynamic information dissemination. Table 1 lists the notation used in the system model. A broadcasting server interacts with clients over the wireless medium. The downward bandwidth is partitioned into a series of interleaved data slots of equal sizes. Suppose that the slotted time model is employed in this model. It takes one time slot to disseminate each data item. Thus, the terms of data slot and time slot are interchangeable where there is no ambiguity. Data slots are further classified as the push or pull mode. Logically, the push slots are viewed as a push channel and those pull ones are as a pull channel.
The broadcasting server disseminates two sorts of data: static data and dynamic data. The former is mapped to the original data broadcast scenario, where the server maintains a broadcast database D and applies a broadcast program P to cyclically deliver static data items, as particularly addressed in the transaction-based and the broadcast disk models. In contrast, the latter is generated dynamically
from information broadcast services, denoted as Is. Each Ii
has a dynamic item production rate, denoted as i, within a
time interval of a broadcast cycle, denoted as L. Contrarily, data are statically kept in the broadcast database in the traditional data broadcast models.
As for the respective access commonality of dynamic and static data, we devise a uniform building block, called as a (service) group, for the structural representation in this model. Two types of groups are specified in the GID mechanism: virtual group and actual group, denoted as
V Gi and AGi, respectively. Clients, who have the same
interest in a static item, virtually form a group; the server
broadcasts a static item to clients in a V Giat the same time.
Note that we use the word “virtual” because there is no
membership among clients in reality. Hence, a V Gi is
backward compatible with the traditional data broadcast
scenario. In contrast, an AGi includes a number of clients
who have subscribed the same information broadcast
service Ii. In such a setting, the group classification policy
is the relative group popularity, denoted as GPi, as will be
defined later. Accordingly, a group is classified into a hot/ cold group; correspondingly, data items in a group are hot/ cold. Without loss of generality, given a specific broadcast bandwidth (M data slots), the server maintains a flat broadcast program to disseminate hot group-specific items periodically. Items in cold groups are delivered by the request-response way in response to their members’ pull requests from the uplink channels. Several primitives in this model are as follows:
1. Each group item is self-identified and read-only.
2. An item slot can alternatively be switched in the pull
or push mode.
3. A client must at least belong to an actual/virtual
group, and its access behavior should be regulated by its group membership.
4. The broadcasting server allocates a slot quota, i.e., a
number of data slots, to a group if its group popularity is higher relatively.
This model is similar to those in [4], [12], [20], [36], while having a graceful extension for the design of the GID framework.
3.2 Design Abstraction
Fig. 3 illustrates the cyclic flowchart of the GID mechanism with the periodicity of data broadcasting. In the beginning of each broadcast cycle, the server initially groups the access commonality of static and dynamic data. Each group is assigned with a broadcast slot quota for disseminating the group-specific data items. During each cycle, the server analyzes the dynamic traffic patterns and the use of slot quotas of all groups. Accordingly, in the end of the broadcast cycle, the server manipulates the group associa-tion and the group popularity of each group. Moreover, the hot and the cold group sets are updated by the group classification. In light of the group schedule priority, the server executes the group replacement, i.e., a series of
group demotion and promotion operations, to adjust the broadcast program for the next broadcast cycle. Due to the variety of item production rate, we devise the loan-based feedback control for the dynamic slot allocation as will be described in the next section.
3.3 Group Association
In the GID mechanism, the actual group and the virtual group are defined to represent the dynamic and the static information broadcast services with a uniform representa-tion structure.
. Actual Group: In the subscription-based dynamic
information scenarios, the broadcasting server groups the clients who have subscribed the same
information service Ii as an AGi. Let kAGik denote
the number of clients in AGi. The value of kAGik
also indicates the amount of message duplications in
response to an item di generated dynamically by Ii.
Provided that Iihas a dynamic item production rate
i, the message traffic for AGi is kAGik i within
the time interval of a broadcast cycle L.
. Virtual Group: To backward support the original
functionality of data broadcasting, each static item in the database is mapped to a virtual information service which always generates the same item. The server assumes that a client with access interest in a
static item dijoins V Gi. Thus, V Giis a special case in
the framework with i¼ 1, and the number of
clients in V Gi, denoted as kV Gik, is the access
frequency for item di.
It is assumed that the initial value of i in AGi and the
initial access frequency of a static item in V Giare obtained
from the prior knowledge of traffic patterns. Nevertheless, in light of the LSAFC technique, as will be mentioned in Section 4, the server is able to adapt the group association responding to dynamic changes of these two factors.
3.4 Group Popularity
This section formulates the quantification of the group popularity and then presents the ways to consolidate the quantification due to dynamic traffic changes.
TABLE 1
3.4.1 Quantification of Group Popularity
Whereas the use of data broadcasting has the significant provision of scalability, the message traffic should be minimized when the server disseminates static and dy-namic data. In light of Definitions 1 and 2, the GID mechanism quantitates the group popularity, denoted as
GPi, in the period of a broadcast cycle.
Definition 1.The GPiof an AGiis the message traffic by Ii, i.e.,
kAGik i.
Definition 2.The GPiof a V Giis the number of clients in V Gi,
i.e., kV Gik.
Note that the item access frequency is a preemptive factor in the calculation of message traffic. In the conven-tional data broadcast model, the access frequency of a static item is directly proportional to the number of interested clients during the time interval of a broadcast cycle, i.e.,
kV Gik in a V Gi. In contrast, the message traffic of an AGiis
the multiple of the number of clients in an AGi and
dynamic item production rate.
3.4.2 Consolidation of Group Popularity
In practice, the measure of dynamic GPi of a V Gi is
difficult. The work in [20] has proposed the Selective Deferment and Reflection (SDR) technique to estimate dynamic item access frequency by reflectively exploiting the notion of client impatience. Briefly, provided that the clients have a mean patience (! time slots)\ in waiting for a
broadcast item di in V Gi, the estimated mean access
frequency of item di in L can be obtained with a reflective
base dj as GPi ¼ kV Gik ¼ i¼ i L e L2 2! j rj ; ð1Þ
where iis the estimated mean access frequency of item di,
and riis the mean number of impatient requests for item di.
On the other hand, the consolidation of GPi of an AGi
uses the mean value of kAGik i. Since kAGik and iare
various during a broadcast cycle, the server calculates the
mean GPi as GPi¼ P 1kLkAGiðkÞk i L ; Gi 2 U A; ð2Þ
where k indicates the kth slot within a broadcast cycle L.
Note that the number of subscribed clients in an AGi is
known, transparently, by the broadcasting server on behalf
of the information broadcast service Ii. In addition, the
LSAFC technique performs dynamic slot allocation during each broadcast cycle and functionally provides the
calcula-tion of dynamic item produccalcula-tion i. Therefore, the server is
able to consolidate GPi of each AGi.
3.5 Group Classification and Broadcast
Program Generation
The relative group popularity is the basic policy for group classification and broadcast program generation. Without loss of generality, the higher the group popularity, the hotter a group will be. Given with M broadcast bandwidth slots, the server iteratively selects the group of the largest group popularity into the broadcast program P . Each group
AGi/V Gi in P is assigned a number of data slots, i.e., the
slot quota, denoted as Si. Note that, for an AGi, the value of
Si is equal to dynamic item production rate i as measured
in the last cycle; in contrast, the slot quota of a V Gi owns
one slot. Accordingly, the iterative selection will continue if and only if the aggregate of allocated slots to all groups in P is less than or equal to M. Thus, we have the length of a broadcast cycle L as
L¼X
8AGiin PSiþ
X
8V Giin P1 M: ð3Þ
In this context, we define that a group is a hot group if it is scheduled in P ; otherwise, that group is a cold group. As
such, all groups in P form the hot group set UH, and others
form a cold group UC. Note that the server will not continue
to look for a group of lower group popularity to digest the remanent bandwidth slots, i.e., M L slots, so as to ensure
that all groups in P are relatively hotter than others in UC. In
addition, the schedule positions of the group-specific items in P cannot be predetermined as a result of dynamic data contents. For that reason, the study in this paper arranges a flat broadcast program, where the scheduled position of any group-specific item is flexible and dependent on the input sequence of dynamic data, in comparison with the static schedule optimization in other previous works.
3.6 Group Schedule Priority
Notwithstanding the server schedules the broadcast pro-gram according to the relative group popularity, it is a critical situation that multiple groups have the same group popularity. For the sake of efficacy and robustness of the broadcast program, two auxiliary definitions, Definitions 3 and 4, are given for the server to determine the relative schedule priority.
Definition 3.Given two groups Ga and Gb with GPa¼ GPb,
Gahas a higher priority than Gbif Gais a virtual group and
Gb is an actual group.
Definition 4.Given two actual groups AGa and AGb with
GPa¼ kAGak a¼ GPb¼ kAGbk b;
AGahas a higher priority than AGbif the factor difference a
is smaller than b, where the factor difference is expressed as
i¼ j k AGik ij.
We apply Definition 3 because the group popularity of a virtual group is more stable than that of an actual group.
Explicitly, referring to Definition 2, the GPi of a V Gi is
mainly dominated by the various kV Gik with a constant
production rate i¼ 1. In contrast, the GPi of an AGi is
more sensitive to the dynamic changes of kAGik and i, as
specified in Definition 1. In addition, Definition 4 states that when two actual groups have the same group popularity,
the one with a smaller i is less sensitive and is selected
earlier. In practice, it is believed that, for an AGi with a
larger i, the increase or decrease in the value of ior kAGik
can lead to a larger variation of GPi. For example, suppose
that AG1 and AG2 have GP1¼ GP2, where kAG1k ¼ 6,
1¼ 4, kAG2k ¼ 12, and 2¼ 2. AG1has a higher schedule
priority than AG2because 1¼ 2 < 2¼ 10 and the change
of GP1is smaller than that of GP2in the case that both of 1
and 2 are increased by one item.
In the GID mechanism, the mean factor difference is used to consolidate the determination of the group schedule priority. After each broadcast cycle, the server has the values of AGi k k ¼ P 1kLkAGiðkÞk L ; i, and i¼ kAGik i of each AGi. Consequently,
according to (1) and (2), and the schedule priority, the server is able to cyclically adjust the group classification and the broadcast program by (3).
4
D
ESIGNo
FLSAFC T
ECHNIQUEThis section devises an online loan-based slot allocation and feedback control (LSAFC) technique to enhance the perfor-mance of the GID mechanism. Section 4.1 presents the procedure of the LSAFC technique. Section 4.2 presents three loaning policies for dynamic slot allocation. Section 4.3 describes the feedback control for the GID mechanism. The algorithmic procedure of the GID+LSAFC mechanism is given in Section 4.4.
4.1 Procedure of the LSAFC Technique
Even though the GID mechanism adjusts the slot allocation periodically in the beginning of each broadcast cycle, the slot allocation is still vulnerable during data broadcasting, while the traffic changes dynamically. Particularly, the slot quota assigned to a group can be excess or scarce. Hence, the design of the LSAFC technique complements the GID mechanism functionally to improve the use of broad-cast bandwidth and the performance.
Fig. 4 illustrates the procedure of the LSAFC technique. In the GID framework, as mentioned in Section 3, each group in the broadcast program is assigned with a slot quota in the start of each broadcast cycle. After that, the
server performs dynamic slot allocation by allowing the loan of slots among groups during the broadcast cycle. The loan-based slot allocation is processed whenever some slots are available before the end of the broadcast cycle. Particularly, a group can loan slots from other groups to disseminate additional data items, in light of the specific loaning policy, once a group has exhausted its slot quota. However, it is possible that a group cannot get a loan of slots because the aggregate of dynamic data items exceeds the total slot quota reserved for the broadcast program. To resolve it, the server will deliver the excess items either by the pull way or by the push way with slot preemption in the next broadcast cycle. In addition, in the end of a broadcast cycle, the server has the loan-based feedback information for the GID mechanism to control accordingly the adapta-tion of slot allocaadapta-tion and broadcast program generaadapta-tion. Therefore, the LSAFC technique enables the GID mechan-ism to use the broadcast bandwidth efficiently and avoid the performance degradation for dynamic information dissemination.
4.2 Dynamic Slot Allocation
An efficient dynamic slot allocation is significant in the design of a robust LSAFC technique. Since the broadcast traffic is unpredictable, some heuristics are applied to devise the loaning policies. Accordingly, we present three loaning policies: the sensitive loan, the insensitive loan, and the greedy loan, for the server to determine the slot loaner. Note that dynamic slot allocation enables the server to tolerate the dynamic data generation.
Let a variable t present the time moment during a broadcast cycle with 0 t L. The time interval ½0; t means the elapsed time from the beginning of the broadcast cycle to the current moment. Suppose that the server
receives a dynamic item difrom the information service Iiat
t. The server will take one slot out of the slot quota Siof the
group Gi to broadcast item di. In case the slot quota is used
up, i.e., Si¼ 0, the server will attempt to loan a slot from
another group decided by one of the following policies:
. Sensitive Loan (SL). Let St
i denote the number of
used slots of each group Gi within ½0; t. The
server estimates the total number of slots that will be used by each group within ½0; L as Si0¼ St
iLt. Thus, the estimated number of
remained slots of each group at the moment t ¼
L is given as Si¼ Si S0i¼ Si SitLt.
Accord-ingly, the group with the maximal Si is selected
as the slot loaner.
. Insensitive Loan (IL). The server considers the slot
quota Siof each group as the base and the number of
used slots St
i without regard to the future traffic
within ½t; L. The server decides the group with the
maximal value ofSiSti
Si ¼ 1
St
i
Si as the slot loaner.
. Greedy Loan (GL). The group which has the
maximal number of remained slots, i.e., Si Sti, at
the moment, t is served as the slot loaner.
In the SL policy, for each group, the server analyzes the temporary pattern of slot utilization, and then accordingly estimates the overall slot utilization progressively to the end of the broadcast cycle. Note that the SL policy is responsive to the decrease or increase of dynamic item production rate. Oppositely, the IL policy adopts a conservative premise that the slot utilization of each group will be the same as that in last broadcast cycle. Hence, the IL policy is less susceptible to the bursty traffic change. In addition, we investigate the feasibility of the GL policy without concern of any traffic patterns. These three loaning policies will be examined comparatively in Section 5.
4.3 Feedback Control
According to the loan-based feedback from the dynamic slot allocation, the server is able to control the adaptation in the GID mechanism, thereby maintaining the performance. This section describes the process of the feedback control below.
4.3.1 Direct Feedback
In the end of each broadcast cycle, the server acquires the information of slot utilization during the time period of a
broadcast cycle. For each group Giwith an initial slot quota
Si, the server knows the number of slots lent to other groups
Si , the number of slots borrowed from other groups Siþ, the
number of remained slots S
i , and the number of items kept
in the queue Qi. Thus, the server can learn the amount of
slots taken to deliver all group-specific items of Gi by the
calculation of the direct feedback (DF) method as
DF : Si0¼ Siþ Sþi Si Siþ Qi; ð4Þ
where S0
iis the required slot quota to meet the actual need,
that is, the dynamic item production rate 0
iin this broadcast
cycle. We notice that either Sþi or Siis equal to zero in the
end of every broadcast cycle, hereof depending on the
various 0
i. Particularly, a group will try to borrow slots
from other groups only if it has used up the slot quota. In
case 0
i is larger than Si, we have Siþ 0 and Si¼ 0;
otherwise, we have Siþ¼ 0 and Si 0. In contrast, the
values of Sþi and Si are decided by the specific loaning
policy and the data sequence received by the server. On the other hand, when a broadcast cycle finishes, there may exist
Qi items of group Giin the queue because the aggregate of
dynamic items of all groups possibly exceeds the band-width accommodation for the broadcast program.
Consequently, the DF method is used to measure dynamic item production rate reflectively from the feed-back information. Along with the measure of group popularity in Section 3.4, accordingly, the server is able to further perform the adaptation, as specified in Section 3, for the next broadcast cycle.
4.3.2 Dissemination of Queued Items
Note that subject to the real-time restriction, the server must digest the queued items in some “real-time” fashions in advance of the next broadcast cycle. Based on the hybrid data broadcast model, two ways are considered below:
. Pull: The server delivers the queued items by the
pull way in the end of the current broadcast cycle.
. Push with slot preemption: The server preempts
slots in the next broadcast cycle and broadcasts the queued items by the push way.
However, disseminating queued items gives rise to an uncertain increase or decrease of message traffic. In practice, because the server has no knowledge of the future traffic, it is problematic in the selection of the pull way or the slot preemption. To provide more insight, several experiments are conducted to inspect the performance comparison as will be given in Section 5.
4.4 GID+LSAFC Procedure
In light of Sections 3 and 4, we integrate the procedures of the GID mechanism and the LSAFC technique as follows: Step 1. Input: A group set.
a. The broadcasting server computes the group
popu-larity GPi of each AGi V Gi.
b. According to the relative group popularity, the
server performs group classification.
c. The server generates the broadcast program P by
iteratively selecting a group Gi of the maximal
schedule priority into P . In addition, each group in
P is assigned with a slot quota Si with (3).
Output: a flat broadcast program P , the length of a
broadcast cycle L, the hot group set UH, and the cold group
set UC.
Step 2. Input: P , L, and a queue Q which contains the received data items subsequently.
a. In the time period of L, the server measures the
variations of kAGik, i, and kV Gik.
b. In the time period of L, the server iteratively fetches
an item in Q and broadcasts it.
. The server sequentially broadcasts the item of
group Gi of the maximal group schedule
priority.
. The server decreases Si by one. In case Gi has
exhausted its slot quota, i.e., Si¼ 0, Gi will loan
a slot from another group in accordance with the GL/SL/IL loaning policy except that all groups have used up their slot quotas.
Step 3. In the end of L,
a. The server checks the queue Q and selects either the
pull way or the push way with slot preemption for disseminating the queued items.
b. In accordance with the loan-based feedback
informa-tion, the server uses the direct feedback method to
calculate the mean group popularity GPiof each Gi.
c. The server repeats Step 1 to adapt P , L, UH, UC, and
Si of each Gi in P .
After each broadcast cycle, the server repeats this procedure to adapt group popularity, group classification,
and broadcast program. Note that each Giin P has Si slots
corresponding to its i. When the remaining slots are
will never look for a group that has lower schedule priority and a smaller item production rate to take those remanent slots. Therefore, the server ensures that all groups in P are
relatively hotter than others in UC.
5
P
ERFORMANCEE
VALUATIONSection 5.1 models the simulation environment. Section 5.2 specifies the performance metrics. Section 5.3 inspects the GID mechanism in comparison with the unicast model. The LSAFC technique with various loaning policies is examined in Section 5.4. In Section 5.5, several experiments are conducted to demonstrate the efficacy of the GID+LSAFC mechanism. Finally, Section 5.6 summarizes significant remarks.
5.1 Simulation Environment
Table 2 lists the simulation parameters description. In accordance with Definitions 1 and 2, the aggregate of the
traffic workload in a time unit is the summation of WA¼
P
AGikAGik i and WV ¼
P
V GikV Gik. Given with b
broadcast slots, the server generates the broadcast program
by the relative group popularity. Because an AGi has
GPi ¼ AGk ik i, the simulator tunes the ratiokAGiik in the
range of 1
4;
1 2; 1; 2; 4
to examine the influence of various traffic factors.
5.1.1 Access Frequency Distribution
WA and WV are generated by a Zipf distribution with a
skew coefficient [42], expressed as pi¼ 1 i =X 1iL 1 i ;
which is used to model the skewed access frequency distribution [7], [17], [20], [29], [39]. The access frequency
ifor an item diis pi . The distribution becomes skewer as
increases and reduces to a uniform distribution as ¼ 0.
5.1.2 Dynamic Traffic Generation
We describe the generations of dynamic workload and dynamic access pattern below. A joint change of workload and access pattern is also considered in this simulation:
. Dynamic workload: The increase or the decrease of
dynamic WAis mapped to the geometric series with
a rate 1 < < 1, that is,
WA¼ fWA0ð1 þ Þ
0; W
A0ð1 þ Þ
1 ; . . .g:
In addition, the dynamic WV is generated similarly.
. Dynamic access pattern: Given with a destined skew
coefficient 2, the skew access pattern in the ith run is
generated by the Zipf distribution with i¼ 1þ
i
Rð2 1Þ and 1 i R.
5.2 Performance Metrics
This section specifies the server-centric performance metrics to evaluate the GID mechanism and the LSAFC technique. Note that the user-centric metrics, e.g., mean access time, adopted usually in the traditional model, are not applicable in dynamic data dissemination contexts, while data are produced dynamically. In contrast, the user-central metrics are used to evaluate the static optimization of the indexing and the scheduling techniques in the traditional model where the broadcast schedule or data contents are static and predetermined in advance of each broadcast cycle.
5.2.1 Measure of Dynamic Message Traffic
Based on an extended hybrid data broadcast model, the GID+LSAFC mechanism broadcasts dynamic data items of hot groups by the push way, delivers dynamic items of cold groups by the pull way, and disseminates the queued items by either the pull way or the push way with slot preemption. Accordingly, the measure of dynamic message traffic within a cycle time includes the broadcast traffic of hot actual groups and virtual groups in P , the pull traffic of cold actual groups and virtual groups, and the extra traffic caused by delivering queued items in the end of a broadcast cycle. The respective mathematical expressions are given below.
GID+LSAFC and the pull way for the queued items:
T ¼ X 8AGi2UH Si0 Qi 1 þ Qi AGk ik þ X 8V Gi2UH 1þ X 8AGi2UC i AGk ik ð Þ þ X 8V Gi2UC V Gi k k:
GID+LSAFC and the push way for the queued items with slot preemption: T ¼ X 8AGi2UH Si0þ X 8V Gi2UH 1þ X 8AGi2UC i AGk ik ð Þþ X 8V Gi2UC V Gi k k þ T ; TABLE 2
where T is the possible decrease/increase of message traffic caused by the slot preemption. Particularly, the slot preemption decreases the number of broadcast slots for the broadcast program in the next broadcast cycle; nevertheless, the message traffic may decrease or increase due to the dynamic traffic changes. Hence, the relative performance is examined in the simulation.
5.2.2 Percent of Loaned Slots
Under the same traffic conditions, the percent of loaned slots is used to assess the relative superiority among the SL, IL, and GL loaning policies.
Percent of loaned slots¼
the number of loaned slots
the length of the broadcast cycle 100%:
Note that the percent of loaned slots indicates the robustness of a loaning policy and the overhead of slot maintenance and computation. Explicitly, the computation overhead is the cost to determine a slot loaner, and the maintenance of slot allocation causes extra overhead of state space.
5.3 Evaluation of Basic GID
In this section, we conduct several experiments to evaluate the scalability of the GID mechanism in terms of the sensitivity to dynamic traffic factors: item production rate, workload and skew access pattern, as considered in the simulation.
In Fig. 5a, the linear increase of message traffic by unicasting dynamic items obviously suggests the super-iority of the GID mechanism. With the increment of broadcast bandwidth, the broadcast program maintains a larger number of groups and delivers the group-specific data on the broadcast channel. Correspondingly, the tendency of the increase of message traffic is diminished. Eventually, the message traffic has no changes when the broadcast bandwidth is enough to accommodate dynamic data of all groups. In addition, Fig. 5b presents the
experimental results regarding the sensitivity to the skew access pattern. With a skew access pattern, most workload congregates in few groups of relatively larger popularity. The message traffic can be reduced rapidly if the available bandwidth is enough for broadcasting these groups. It is also depicted in Fig. 5b that the skewer the access pattern, the lower the message traffic.
Further, the GID mechanism is examined under dynamic
workloads which are generated with various ratios of WA
WV
and i
AGi
k k. Fig. 6 shows the measured results of group
classification. In Fig. 6a, in case that i is constant as the
value in the initial workload, the number of hot virtual groups in the broadcast program decreases when the
workload WA is increased by the incremental ratio of WWAV.
Oppositely, the number of hot actual groups increases slightly. On the other hand, as illustrated in Fig. 6b, in case
of a constant AGk ik, both of the respective amounts of hot
actual and hot virtual groups decrease when the ratio ofWA
WV
increases. This phenomenon is explained as follows: Most
workload belongs to actual groups, and each AGi needs a
larger slot quota corresponding to the increase of i. In
contrast, there are few virtual groups of higher group popularity in this case. Therefore, the total amount of hot
groups decreases in response to the incremental WA
WV, but
increases under the decrement ofWA
WV.
5.4 Evaluation of LSAFC with Various
Loaning Policies
Several experiments have been conducted to synthetically evaluate the LSAFC technique with different loaning policies in terms of the percent of loaned slots. The numerical results are listed in Table 3.
We notice that using prior traffic patterns to generate a reliable broadcast program, as assumed commonly, is paradoxical in dynamic data dissemination environments. It is showed in Table 3 that the IL policy has the poorest performance. Particularly, when the tendency of traffic
Fig. 5. (a) Various i
AGi
k k(experimental baseline: ¼ 10, AG ¼ V G ¼ 100, and ¼ 0:8), and ratio r : kAGk. (b) skew access pattern (experimental
changes and the slot utilization are similar to the past patterns, the server can allocate each hot group a slot quota near to the actual need within a broadcast cycle. Due to the unpredictable input sequence of dynamic data, however, dynamic slot allocation with the IL policy is inefficient. In contrast, the SL policy takes the temporary traffic patterns into consideration and attains the best performance. On the other hand, we deliberate that the GL policy is prominent with its simplicity and less computation overhead, although the SL policy is better than the GL in a degree of 1 2%. Therefore, the GL policy is considered when the broadcast cycle is lengthy, or when the overhead of slot allocation is critical.
Furthermore, as the comparison with parts 1, 2, and 6 in Table 3, the amount of loaned slots decreases under a higher skew traffic ( ¼ 0:8 > 0:6 > 0:2). When the skew access pattern is stable, some slots may remain, if the
workload decreases, as the comparison between parts 2 and 3, and parts 5 and 6. In case the access pattern becomes skewer drastically, the percent is decreased with more available bandwidth slots. Since most workload congre-gates to few hot groups, other groups have lower possibility to loan slots. In addition, while scheduling the broadcast program, the server ensures that the groups in the broad-cast program are relatively hotter than other cold groups. A group of lower schedule priority cannot be taken to digest the remanent b L slots. Consequently, the length of the broadcast cycle L is slightly smaller or equal to b as shown in Table 3.
5.5 Evaluation of GID+LSAFC
This section presents the investigation on the GID+LSAFC under dynamic workload and dynamic skew access pattern. The measured message traffic is compared with the local optimum, that is, the comparison baseline which is available only if the server could perceive dynamic traffic patterns in the beginning of each broadcast cycle.
5.5.1 Dynamic Workload
Fig. 7 depicts the experimental results under the increas-ing and the decreasincreas-ing workloads. Observe that the GID+LSAFC mechanism is able to reduce the message traffic prominently. According to the loan-based feedback, the GID mechanism is able to perform the adaptation of group popularity, group classification, broadcast program, and slot allocation in the end of the broadcast cycle. Regarding the delivery of excess items, it is shown in Fig. 7a that the pull way has the better performance when the workload increases. In contrast, as depicted in Fig. 7b, the push way with slot preemption is the best when the workload decreases; noteworthily, it can minimize the message traffic to be lower than the local optimum based on the local traffic pattern within the current cycle. Explicitly, the decreasing workload will cause the number of hot items to be smaller than the total slot quota in the next broadcast. Thus, slot preemption will not result in the
Fig. 6. (a)WA
WV versus constant , and constant item production rate. (b)
WA
WV versus constantkAGik (experimental baseline:
i
AGi
k k¼ 1, ¼ 6,
AG¼ V G ¼ 200, and ¼ 0:6), and constant number of clients.
TABLE 3
The Measured Results by the LSAFC Technique with the GL, SL, and IL Loaning Policies (Experimental Baseline:
AG¼ V G ¼ 200, i
AGi
increase of message traffic. In addition, as plotted in Fig. 7a, when the slot preemption is used, the message traffic in the GID mechanism is close to that in the GID+LSAFC mechanism. However, it is noted that the former signifi-cantly incurs about 4 5 times the buffer size required for temporarily keeping the excess items in the latter. Further-more, using slot preemption to broadcast queued items is less susceptible to dynamic workload and may not lead to a drastic increase of message traffic. In contrast, the message traffic by unicasting the queued items is critical and related to the number of queued items and the item popularity. Consequently, the GID+LSAFC mechanism is able to get the best performance, while the workload increases. Otherwise, the GID+LSAFC mechanism with the slot preemption is more profitable under a decreasing workload.
5.5.2 Dynamic Skew Access Pattern
Several experiments are conducted to evaluate the relative performance under dynamic skew access patterns. Fig. 8 illustrates the experimental results in the increment and the decrement of skew coefficients. Generally speaking, more
workload congregates in fewer groups with the increment of skew coefficient. The overall message traffic is reduced by allocating broadcast slots first to the relatively hotter groups. In addition, there are several observations regard-ing the delivery of excess items as follows: First, as depicted in Fig. 8a, the extra message traffic caused by delivering the excess queued items reduces gradually in reverse propor-tion to the incremental skew coefficient. Deductively, there may be no extra pull message traffic in case of more bandwidth or a drastic increment of skew coefficient since all excess items can be satisfied by loaning slots alterna-tively. Second, the push way with slot preemption has a substantial reduction of message traffic in comparison with the pull way. Particularly, given a broadcast program, more slots will remain when the skew access pattern becomes skewer. Hence, the slot preemption in the next cycle will not much affect the dissemination of dynamic data in the next broadcast program. On the other hand, as compared with the case under dynamic workload, the variation of message traffic is more sensitive to the incremental skew coefficient due to the higher frequency of dynamic slot allocation. In
Fig. 7. Comparison under dynamic workload (experimental baseline: R¼ 10, b ¼ 200, ¼ 20, AG ¼ V G ¼ 200, i
AGi
k k¼ 1, and ¼ 0:6).
Fig. 8. Comparison under dynamic skew access pattern (experimental baseline: R¼ 10, b ¼ 200, ¼ 20, AG ¼ V G ¼ 200, and i
AGi
contrast, in the case of the decremental skew coefficient, the slot allocation based on the feedback in the last broadcast cycle is sufficient to sustain dynamic item production rates. Remarkably, the GID+LSAFC mechanism with slot pre-emption is able to reduce the message traffic prominently under dynamic skew access patterns. Further, the pull way is able to attain similar performance to that of the push way when the access pattern becomes smoother.
5.6 Summary and Discussion
To make a synthetic comparison, we further investigate the relative performance under the joint changes of dynamic workload and skew access pattern. In light of extensive experiments in the simulation, several remarks are made below.
The GID+LSAFC mechanism has demonstrated the scalability for dynamic data dissemination in terms of message traffic. We have evaluated the design of the GID mechanism and the LSAFC technique against the
performance impacts of the system-designed factors with various traffic conditions. The efficiency and robustness of the GL, IL, and SL loaning policies are comparatively investigated in terms of the percent of loaned slots and the overhead of slot maintenance. It is shown that the dynamic slot allocation is more efficient, while the temporary traffic patterns during the cyclic broadcast is considered, instead of the past traffic patterns.
As shown by the experimental results, remarkably, the GID+LSAFC mechanism achieves the best performance by using the pull way to disseminate queued items, while dynamic workload increases. In contrast, the GID+LSAFC mechanism with slot preemption for broadcasting queued items is more profitable in response to the incremental skew of dynamic access pattern. Furthermore, the performance difference between the GID mechanism and the GID+LSAFC mechanism, as illustrated in Fig. 9, is more discernible under a joint change of dynamic workload and access pattern. Comparatively, the LSAFC technique enables the GID mechanism to perform
Fig. 9. Comparison under various joint changes of dynamic workload and skew access pattern (experimental baseline: ¼ 25, AG ¼ V G ¼ 200, and
i
AGi
k k¼ 1). (a) W ¼ þ8%, theta ¼ 0:4 to 1.2, and b ¼ 200. (b) W ¼ 8%, theta ¼ 0:4 to 1.2, and b ¼ 200. (c) W ¼ þ8%, theta ¼ 1:2 to 0.4, and
dynamic slot allocation efficiently for broadcasting hot items as soon as possible. The message traffic caused by delivering excess queued items will thus be reduced. Therefore, the above observations show the advantages and efficacy of the GID+LSAFC mechanism.
6
C
ONCLUSIONSThe work in this paper has devised the GID+LSAFC mechanism, essentially an adaptive information dissemina-tion framework, by exploiting the potential of data broad-casting. The GID mechanism not only supports dynamic information dissemination, but is also backward compatible with the conventional data broadcast scheme. Considering both of the client-oriented and the server-oriented traffic factors, we have further designed an online LSAFC technique to sustain dynamic traffic changes. Based on the LSAFC technique, the GID mechanism is able to perform the adaptations of broadcast slot allocation, group associa-tion, popularity and classificaassocia-tion, and broadcast program. Extensive simulations have been conducted to evaluate the performance of the GID+LSAFC mechanism. The experi-mental results have shown that the GID mechanism is very scalable and attains a substantial reduction of dynamic message traffic. In addition, the LSAFC technique is able to complement the GID mechanism functionally, therefore enhancing the performance and making the GID mechan-ism more efficient and robust.
A
CKNOWLEDGMENTSThe authors are supported in part by the Ministry of Education Project No. 89-E-FA06-2-4, and the National Science Council Project No. NSC 92-2213-E-002-001 and NSC 92-2213-E-002-010, Taiwan, Republic of China.
R
EFERENCES[1] The 3rd Generation Partnership Project, http://www.3gpp.org, 2003.
[2] S. Acharya, R. Alonso, M.J. Franklin, and S.B. Zdonik, “Broadcast Disks: Data Management for Asymmetric Communications Environments,” Proc. ACM Int’l Conf. Management of Data, pp. 199-210, May 1995.
[3] S. Acharya, M.J. Franklin, and S.B. Zdonik, “Prefetching from Broadcast Disks,” Proc. 12th IEEE Int’l Conf. Data Eng., pp. 276-285, Feb. 1996.
[4] S. Acharya, M.J. Franklin, and S.B. Zdonik, “Balancing Push and Pull for Data Broadcast,” Proc. ACM Int’l Conf. Management of Data, pp. 183-194, May 1997.
[5] S. Acharya and S. Muthukrishnan, “Scheduling On-Demand Broadcasts: New Metrics and Algorithms,” Proc. ACM MobiCom, pp. 43-54, 1998.
[6] D. Aksoy and M. Franklin, “Rxw: A Scheduling Approach for Large-Scale On-Demand Data Broadcast,” IEEE/ACM Trans. Networking, vol. 7, no. 6, pp. 846-860, 1999.
[7] M.H. Ammar and J.W. Wong, “On the Optimality of Cyclic Transmission in Teletext Systems,” IEEE Trans. Comm., vol. 11, pp. 1159-1170, 1987.
[8] A. Bar-Noy, R. Bhatia, J. Naor, and B. Schiber, “Minimizing Service and Operation Costs of Periodic Scheduling,” Proc. Ninth ACM-SIAM Symp. Discrete Algorithms (SODA ’98), pp. 11-20, Jan. 1998.
[9] A. Bar-Noy, J. Naor, and B. Schiber, “Pushing Dependent Data in Clients-Providers-Servers Systems,” Proc. ACM MobiCom ’00, pp. 222-230, 2000.
[10] D. Barbara, “Mobile Computing and Database—A Survey,” IEEE Trans. Knowledge and Data Eng., vol. 11, no. 1, pp. 108-117 Jan./Feb. 1999.
[11] T. Bowen et al., “The Datacycle Architecture,” Comm. ACM, vol. 35, no. 12, pp. 71-81, 1992.
[12] J. Cai and K. -L. Tan, “Tuning Integrated Dissemination-Based Information Systems,” Data & Knowledge Eng., vol. 30, no. 1, pp. 1-21, 1999.
[13] Y.-I. Chang and C.-N. Yang, “A Complementary Approach to Data Broadcasting in Mobile Information Systems,” Data & Knowledge Eng., vol. 40, no. 2, pp. 181-194, Feb. 2002.
[14] M.-S. Chen, P.S. Yu, and K.-L. Wu, “Optimizing Index Allocation for Sequential Data Broadcasting in Wireless Mobile Computing,” IEEE Trans. Knowledge and Data Eng., vol. 15, no. 1, Jan./Feb. 2003. [15] R. Deolasee, A. Kathar, A. Panchbudhe, K. Ramamritham, and P. Shenoy, “Dissemination of Dynamic Data,” Proc ACM SIGMOD ’01, p. 599, May 2001.
[16] H. Dykeman, M. Ammar, and J. Wong, “Scheduling Algorithms for Videotex Systems under Broadcast Delivery,” Proc. IEEE Int’l Conf. Comm., pp. 1847-1851, 1986.
[17] M.J. Franklin and S.B. Zdonik, “Dissemination-Based Information Systems,” Data Eng. Bull., vol. 19, no. 3, pp. 20-30, 1996. [18] S. Hameed and N.H. Vaidya, “ Efficient Algorithms for
Schedul-ing Data Broadcast,” ACM/Baltzer Wireless Networks, vol. 5, no. 3, pp. 183-193, 1999.
[19] C.-L. Hu and M.-S. Chen, “Adaptive Balanced Hybrid Data Delivery for Multi-Channel Data Broadcast,” Proc. IEEE Int’l Conf. Comm., Apr. 2002.
[20] C.-L. Hu and M.-S. Chen, “Dynamic Data Broadcasting with Traffic Awareness,” Proc. 22nd IEEE Int’l Conf. Distributed Computing Systems, July 2002.
[21] Q. Hu, D.L. Lee, and W.-C. Lee, “Dynamic Data Delivery in Wireless Communication Environments,” Proc. ER ’98 Workshops Mobile Data Access, pp. 218-229, 1998.
[22] Q. Hu, W.-C. Lee, and D.L. Lee, “Indexing Techniques for Wireless Data Broadcast under Data Clustering and Scheduling,” Proc. Eighth Int’l Conf. Information and Knowledge Management, 1999. [23] T. Imielinski and B.R. Badrinath, “Data Management for Mobile
Computing,” SIGMOD Record, vol. 22, no. 1, pp. 34-39, 1993. [24] T. Imielinski and S. Viswanathan, “Adaptive Wireless Information
Systems,” Proc. ACM Special Interest Group on DataBase Systems, pp. 19-41, 1994.
[25] T. Imielinski, S. Viswanathan, and B.R. Badrinath, “Data on Air: Organization and Access,” IEEE Trans. Knowledge and Data Eng., vol. 9, no. 3, pp. 353-372, May/June 1997.
[26] Y.-B. Lin, “A Multicast Mechanism for Mobile Networks,” IEEE Comm. Letter, vol. 5, no. 11, pp. 450-452, 2001.
[27] S. Paul, Multicasting on the Internet and Its Applications. Boston: Kluwer Academic Publishers, 1998.
[28] W.-C. Peng and M.-S. Chen, “Dynamic Generation of Data Broadcasting Programs for a Broadcast Disk Array in a Mobile Computing Environment,” Proc. Ninth ACM Int’l Conf. Information and Knowledge Management, Nov. 2000.
[29] W.-C. Peng and M.-S. Chen, “Efficient Channel Allocation Tree Generation for Data Broadcasting in a Mobile Computing Environment,” Wireless Networks, vol. 9, no. 2, pp. 117-129, 2003. [30] E. Pitoura and R. Chrusanthis, “Scalable Processing of Read-Only
Transactions in Broadcast Push,” Proc. IEEE Int’l Conf. Distributed Computing Systems, Sept. 1999.
[31] E. Pitoura and R. Chrusanthis, “Multiversion Data Broadcast,” IEEE Trans. Computers, vol. 51, no. 10, pp. 1224-1230, Oct. 2002. [32] E. Pitoura, R. Chrusanthis, and K. Ramamritham, “Characterizing
the Temporal and Semantic Coherency of Broadcast-Based Data Dissemination,” Proc. Ninth Int’l Conf. Database Theory, 2003. [33] A.K. Salkintzis, “A Survey of Mobile Data Networks,” IEEE
Comm. Surveys, vol. 2, no. 3, pp. 2-18, 1999.
[34] N. Schabanel, “The Data Broadcast Problem with Preemption,” Proc. 17th Int’l Symp. Theoretical Computer Science (STACS ’00), pp. 181-192, Feb. 2000.
[35] J. Shanmugasundaram et al., “Efficient Concurrency Control for Broadcast Environment,” Proc. ACM SIGMOD, 1999.
[36] K. Stathatos, N. Roussopoulos, and J.S. Baras, “Adaptive Data Broadcast in Hybrid Networks,” Proc. 23rd Int’l Conf. Very Large Data Bases, Aug. 1997.
[37] C.-J. Su, L. Tassiulas, and V.J. Tsotras, “Broadcast Scheduling for Information Distribution,” ACM/Baltzer Wireless Networks, vol. 5, no. 2, pp. 137-147, 1999.
[38] K.-L. Tan and B.C. Ooi, Data Dissemination in Wireless Computing Environments. Boston: Kluwer Academic Publishers, 2000.
[39] N.H. Vaidya and S. Hameed, “Scheduling Data Broadcast in Asymmetric Communication Environments,” ACM/Baltzer Wire-less Networks, vol. 5, no. 3, pp. 171-182, 1999.
[40] J.W. Wong, “Broadcast Delivery,” Proc. IEEE, vol. 76, no. 12, pp. 1566-1577, Dec. 1988.
[41] K.-L. Wu, P.S. Yu, and M.-S. Chen, “Energy-Efficient Caching for Bandwidth-Limited Wireless Mobile Computing,” Proc. 12th IEEE Int’l Conf. Data Eng., pp. 335-343, Feb. 1996.
[42] G.K. Zipf, Human Behaviour and the Principle of Least Effort. Reading, Mass.: Addison-Wesley, 1949.
Chih-Lin Hu received the BS degree in compu-ter science from National ChengChi University, the MS degree from National ChungHsing University, and the PhD degree in National Taiwan University, respectively, in 1997, 1999, and 2003. He had the honor of receiving the best paper award at the IEEE ICPADS-2000. His research interests include mobile agent technol-ogy, broadcast information management, and active networks.
Ming-Syan Chen received the BS degree in electrical engineering from National Taiwan University, Taipei, Taiwan, and the MS and PhD degrees in computer, information, and control engineering from The University of Michigan, Ann Arbor, in 1985 and 1988, respec-tively. Dr. Chen is currently the chairman of the Graduate Institute of Communication Engineer-ing and also a professor in the Electrical Engineering Department, National Taiwan Uni-versity, Taipei, Taiwan. He was a research staff member at IBM Thomas J. Watson Research Center, Yorktown Heights, New York, from 1988 to 1996. His research interests include database systems, data mining, mobile computing systems, and multimedia networking, and he has published more than 160 papers in his research areas. In addition to serving as a program committee member in many conferences, Dr. Chen served as an associate editor of IEEE Transactions on Knowledge and Data Engineering on data mining and parallel database areas from 1997 to 2001, is on the editorial board of VLDB Journal, Journal of Information Science and Engineering, and Journal of the Chinese Institute of Electrical Engineering, was a distinguished visitor of IEEE Computer Society for Asia-Pacific from 1998 to 2000, and program chair of PAKDD-02 (Pacific Area Knowledge Discovery and Data Mining), program vice-chair of VLDB-2002 (Very Large Data Bases) and ICPP 2003, general chair of the Real-Time Multimedia System Workshop in 2001, program chair of the IEEE ICDCS Workshop on Knowledge Discovery and Data Mining in the World Wide Web in 2000, and program cochair of the International Conference on Mobile Data Management (MDM) in 2003, the International Computer Symposium (ICS) on Computer Networks, Internet and Multimedia in 1998 and 2000, and the ICS on Databases and Software Engineering in 2002. He was a keynote speaker on Web data mining for the International Computer Congress in Hong Kong, 1999, a tutorial speaker on Web data mining for the DASFAA-1999 and on parallel databases for the 11th IEEE International Conference on Data Engineering in 1995, and also a guest coeditor for the IEEE Transactions on Knowledge and Data Engineering on a special issue for data mining in December 1996. He holds, or has applied for, 18 US patents and seven ROC patents in the areas of data mining, Web applications, interactive video playout, video server design, and concurrency and coherency control protocols. He is a recipient of the NSC (National Science Council) Distinguished Research Award in Taiwan and the Outstanding Innova-tion Award from IBM Corporate for his contribuInnova-tion to a major database product, and also received numerous awards for his research, teaching, inventions, and patent applications. He coauthored with his students for their works which received ACM SIGMOD Research Student Award and Long-Term Thesis Award. Dr. Chen is a senior member of IEEE and a member of ACM.
.For more information on this or any computing topic, please visit our Digital Library at http://computer.org/publications/dlib.
![Fig. 1 illustrates a generalized mobile data access frame- frame-work [1]. Due to the lack of an efficient multicast mechanism in the existing mobile data systems, a broad-cast/multicast message incurs a number of message relays and transmission cost with](https://thumb-ap.123doks.com/thumbv2/9libinfo/8774845.213548/3.918.170.689.104.383/illustrates-generalized-efficient-multicast-mechanism-existing-multicast-transmission.webp)
