108 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 11, NO.

1, JANUARY/FEBRUARY 1999

Mobile Computing and DatabasesA Survey

Daniel Barbará

Abstract—The emergence of powerful portable computers, along with advances in wireless communication technologies, has
made mobile computing a reality. Among the applications that are finding their way to the market of mobile computingthose that
involve data managementhold a prominent position. In the past few years, there has been a tremendous surge of research in the
area of data management in mobile computing. This research has produced interesting results in areas such as data dissemination
over limited bandwith channels, location-dependent querying of data, and advanced interfaces for mobile computers. This paper is
an effort to survey these techniques and to classify this research in a few broad areas.

Index Terms—Mobile computing, databases, data dissemination, bandwith, location-dependent queries, interfaces, transaction
management.

——————————F——————————

1 INTRODUCTION

M OBILE computing has become a reality thanks to the

convergence of two technologies: the appearance of
powerful portable computers and the development of fast
(supporting data rates on the order of 10 to 20 Kbits/sec.).
In the latter, the bandwidth is much widerup to 10
Mb/sec. Fixed hosts will communicate over the fixed net-
reliable networks. In the mobile wireless computing envi- work, while mobile units will communicate with other
ronment of the future [33] massive number of low powered hosts (mobile or fixed) via a wireless channel.
computer machines will query databases over the wireless Although a wireless network with mobile clients is es-
communication channels. Mobile clients will often be dis- sentially a distributed system, there are some characteristic
connected for prolonged periods of time due to the battery features that make the system unique and a fertile area of
power saving measures; they will also frequently relocate research. These are:
between different cells and connect to different data servers • Asymmetry in the communications: The bandwidth
at different times. in the downstream direction (servers-to-clients) is
The mobile or nomadic computing environment no much greater than that in the upstream direction
longer requires users to maintain a fixed and universally (clients-to-servers). Moreover, in some systems, the
known position in the network and enables unrestricted clients do not have the capacity to send messages
mobility of the users. Mobility and portability will create to the servers at all. As we mentioned above, even
an entire new class of applications and possibly new the bandwidth in the downstream direction is lim-
massive markets combining personal computing and con- ited: 10 to 20 Kbits/sec. in a cellular network and up
sumer electronics. to 10 Mb/sec. in a wireless LAN.
In this paper, we survey the impact that mobile com- • Frequent disconnections: Mobile clients do not stay
puting has had in the area of data management. We first connected to the network continuously (as fixed hosts
analyze each of the distinctive features of mobile comput- do), but rather users switch their units on and off
ing and how they impact the implementation of databases regularly. Moreover, mobile clients can roam, discon-
for mobile computers, creating new opportunities for re- necting from a cell to connect to another.
search. Then for each of these problems we survey the pa- • Power limitations: Some of the portable units are se-
pers that have appeared in the literature in recent years. verely limited by the amount of energy they can use
Fig. 1 displays the architecture of the general mobile en- before the batteries have to be recharged.
vironment. It consists of two distinct sets of entities: mobile • Screen size: Some of the portable units, such as the
units and fixed hosts. Some of the fixed hosts, called Mobile Personal Digital Assistants, have very small screens.
Support Stations, are augmented with a wireless interface
to communicate with mobile units which are located within Each one of these features has an impact on how data
a radio coverage area called a cell. A cell could be a real cell can be effectively managed in a system with mobile clients.
as in cellular communication network or a wireless local The communication asymmetry, along with the restriction
area network which operates within the area of a building. in power that the mobile units have, make the model
In the former case the bandwidth will be severely limited of broadcasting data to the clients (instead of waiting for
the clients to request specific data items), an attractive
²²²²²²²²²²²²²²²² proposition. This is called data dissemination and is the
• D. Barbará is with the Information and Software Engineering topic of Section 2 in this paper. The limited bandwidth
Department, George Mason University, Fairfax, VA 22030. and the pattern of frequent disconnections have a clear im-
E-mail: [email protected]. pact on how transaction management is implemented and
Manuscript received 20 May 1997; revised 20 Aug. 1998. how data consistency is guaranteed in this environment. This
For information on obtaining reprints of this article, please send e-mail to:
[email protected], and reference IEEECS Log Number 108303 is the subject of Section 3. The roaming of clients through

1041-4347/99/$10.00 © 1999 IEEE

BARBARÁ: MOBILE COMPUTING AND DATABASESA SURVEY 109

Fig. 1. A mobile environment.

different cells opens the possibility of answering queries in to information providing queries that describe their inter-
a way that is dependent on the current position of the cli- est. These queries can be viewed as continuous queries [62].
ent, i.e., implementing location dependent queries. We re- The server also needs to decide whether to send the data
view the work in this area in Section 4. Finally, screen and periodically or aperiodically. Periodic push has the advan-
power limitations have an impact on the kind of interfaces tage of allowing clients to disconnect for certain periods
that can be implemented for data browsing and querying. and still not miss out items (since they will be broadcasted
We cover the work in this field in Section 5. Section 6 offers again after the client is reconnected). Aperiodic dissemina-
some conclusions about the future of this research area. tion on the other hand is a more effective way of using the
bandwidth available. Periodic push has been used in the
past in many systems [8], [69], [27], [19], [31].
2 DATA DISSEMINATION Acharya et al. [5], [1] have proposed the use of a periodic
Data dissemination is the delivery of data from a set of dissemination architecture in the context of mobile systems.
producers to a larger set of clients. This is characterized by They call the architecture Broadcast Disks. Broadcast Disks
an inherent asymmetry in the communications: The band- are novel in two ways:
width in the downstream direction (servers-to-clients) is • They provide a multilevel mechanism that permits
much greater than in the upstream direction (clients-to- data items to be broadcast nonuniformly, so more
servers). (For a survey of data dissemination work see [24].) bandwidth can be allocated to items according to
This model fits very well a mobile system, since mobile cli- their relative importance.
ents are usually unable to transmit data at a very high • Mechanisms for managing the storage in the clients
speed (sometimes this capacity is not even present). How- are devised to tailor caching and prefetching to per-
ever, they can receive data at high rate. form efficiently with the multilevel broadcast.
The model of sending information to a client popula-
Using Broadcast Disks, one can construct a memory hi-
tion, without waiting for specific requests is also called
erarchy in which the highest level contains a few items and
push-based. By pushing data, the servers avoid interrup-
broadcasts them with high frequency while subsequent
tions caused by requests and are allowed to propagate levels contain more and more items and broadcast them
data whose existence would otherwise be ignored by clients. with less and less frequency. In this way, one can establish a
On the other hand, push-based systems have the problem trade-off between access time for high-priority data and
of deciding about the relevance of the data to be broad- that of the low-priority items.
casted to clients. The usefulness of the system depends Fig. 2 shows an example in which two levels are pro-
on the ability to predict the clients needs. A simple so- vided. One can think of this broadcast as being provided by
lution is to allow clients to subscribe to services while pro- two disks: the first containing data item A and the second
viding profiles of their interest [49]. The clients subscribe items B1, B2, and B3. The first disk rotates at a speed three
110 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 11, NO. 1, JANUARY/FEBRUARY 1999

3) Chopping off the slowest part of the broadcast sched-

ule successively larger pieces. This has the effect of in-
Fig. 2. An example of broadcast disks. creasing the available bandwidth for pull. This intro-
duces the disadvantage that if there is still not enough
times as fast as the second. The characterization of broad- bandwidth for pull, performance can degrade consid-
cast programs was presented in [5]. A good program erably since clients will not be able to pull the items
that were not broadcast.
renders approximately fixed interarrival times for each
item and allocates bandwidth to items according to their In [58], the authors discuss an algorithm to rapidly adapt
access probabilities. An effort in optimizing programs is the content of mobile client’s caches based on the misses
described in [16]. that result in on-demand data requests. The algorithm is
Since no broadcast program can perfectly match the developed as a way to effectively support data dissemina-
needs of an individual client, the need to compensate for tion using Direct Broadcast Satellite systems. Their simula-
the mismatches exists. This is done by doing intelligent tion results show that the heuristic of adapting caches
caching and prefetching at the client site. Caching and pre- based solely on information about the miss ratio works
fetching policies for Broadcast Disks have been studied in quite reasonably.
[5], [1]. The mechanisms have been extended to incorporate In [13], the authors present a novel way of sending in-
the effect of updates in [2]. In this last case, the server must validation messages over a limited bandwidth network. By
effectively communicate updates to the client by means of using Invalidation Reports (IRs), the server can notify clients
invalidation or propagation. Acharya et al. [2] studied the about changes in the items that are being cached by them.
way of devising prefetching algorithms that work well un- The IRs are very succinct ways of transmitting this infor-
der update scenarios. mation. For instance, one could send a list of identifiers for
Acharya et al. extend their previous work in [3] to study the items that have changed since the last Invalidation Re-
port was sent. Depending on the technique used, IRs can
the integration of a pull-based and a push-based Broadcast
cause “false negatives,” making a client drop an item from
Disk approach. In a pull-based operation, clients explicitly
its cache when in reality the item is still valid. In return for
request items by sending messages to the server, which in
this anomalous behavior, one gets a shorter report that con-
turn sends the information back to the clients. Combining
sumes less bandwidth. In [13], the authors study the trade-
the two approaches can be beneficial for both server and
offs involved in using different types of Invalidation Re-
clients. Their study assumes two independent data chan-
ports. In [14], the issue of relaxing consistency of the caches
nels, front and backchannel: the frontchannel is used for
is addressed. For instance if the mobile clients are caching
push-based operation, while the back-channel serves as the
stock prices, it may be perfectly acceptable to use values
media for the pull-based operation. The available band- that are not completely up to date, as long as they are
width is shared between these two channels. The study within 0.5 percent of the true prices. This can be accom-
concludes that for the extreme cases of little contention at plished by considering the cached values as quasicopies of
the server or server overload, pure-pull and pure-push so- the values in the server [6]. A quasicopy is a cached value
lutions respectively are the best options. For the other cases, that is allowed to deviate from the true copy in a controlled
the authors propose a hybrid solution, called Interleaved way. Using quasicopies, the IRs can be made even shorter,
Push and Pull (IPP), which allows clients to use the back- thereby saving extra bandwidth for their transmission.
channel to send pull requests for items that do not appear Also, in [14], the authors extend the IRs strategies to dy-
in the Broadcast channel, while the server supports Broad- namically adjust to changes in the query and update rates
cast Disks plus interleaved responses to the pulls in the and the disconnection patterns of the mobile clients. These
frontchannel. IPP suffers of the same bottleneck problems new techniques, called adaptive IRs, take into account the
of any pull approach. To improve the scalability of IPP, the on-demand queries that the clients make when they do not
authors propose three different techniques: find the items in their local caches. In Wu et al. [70], the
1) Adjusting the pull bandwidth, at the expense of the problem of discarding caches in mobile computers that
push bandwidth. This introduces a trade-off between have been disconnected for any period of time (sleepers) is
how fast the queue of pull requests is served versus addressed. In [13], the strategy a mobile unit that discon-
how fast the broadcast delivery is executed. Lower nects is to purge its cache after it comes back in operation.
values (30 percent of the available bandwidth) for the Wu et al. propose a mechanism to decide whether some
pull bandwidth do not produce the best results for items in the cache can still be used by the mobile unit. To do
lightly loaded systems, but they also do not produce so, the mobile units need to contact the server when they
the worse results for saturated systems, and thus they come back from a disconnection and ask if the items in the
are good candidates for dynamic loads. cache are still usable. To decrease uplink traffic, the data-
base is partitioned in groups and elements in the same
2) Providing a pull threshold. With this technique, cli-
group are cached together. In this way, the mobile unit only
ents only request their most expensive misses, i.e.,
asks the server for the validity of specific groups. The
those that will be broadcast in the longest future.
server keeps track of the update history for elements in
The effect of reducing the number of pulls coming
each group, and using an algorithm called Grouping with
from clients is to delay the point at which the server
Cold Update-Set Retention (GCoRe), it determines whether
queue saturates.
or not all the items updated since the mobile computer
BARBARÁ: MOBILE COMPUTING AND DATABASESA SURVEY 111

became disconnected have been included in the last IR. it in N independent pieces (dispersal) in such a way that
If the answer is yes, the cached group can be kept recombining any m of such pieces (n ≤ N) is sufficient to
(since the mobile unit can hear about the updates in the retrieve F (reconstruction). AIDA uses IDA to adaptively
IR). Otherwise, the cached group should be discarded decide the amount of redundancy used in the broadcast.
(there is not enough information in the IR to determine Namely, the number of pieces transmitted n is allowed to
which elements to keep). Using simulation, the authors de- vary from m (no redundancy) to N (maximum redun-
termine that as the group size increases, the uplink traffic dancy). Fig. 3, taken from [17], illustrates how two files A
decreases but the downlink traffic increases. The total and B are broadcast using IDA. File A is dispersed into 10
bandwidth usage first decreases and then increases sug- blocks, of which any five blocks are enough to reconstruct
gesting an optimal operating point. it. For file B the algorithm uses N = 6, m = 3. The pieces are
Liu and Maguire [30] use IRs to propose a mobility- broadcast periodically, using a program as shown in Fig. 4.
aware dynamic database caching scheme for mobile com- There are two “periods” in that program: the broadcast pe-
puting. The IR techniques are combined with knowledge of riod, which extends for eight units and, in the absence of
the mobile user’s behavior to broadcast IRs only within the errors is enough to recover both files; and the program data
user’s mobility area. Using a performance model, the cycle which broadcasts all the pieces from all the dispersed
authors claim the scheme can reduce the system cost by files in the disk. Bestravos finds upper bounds for worst-
more than 87 percent when compared to a fully replicated case delays given the number of transmission errors.
database system. If the data is just broadcast without the any form of
The previous Broadcast Disks protocols are not designed directory, clients have to tune to the channel continuous-
with real-time demands on mind. In real-time environ- ly until the required data items are downloaded. This
ments, minimizing the average latency time is no longer forces clients to be in active mode for a long time, consum-
the main criteria for performance. Rather, guaranteeing ing a lot of power, which is already scarce. By providing a
that time constraints are met is the important concern. directory, the server enables the ability of tuning and that
Also, in the previous Broadcast Disks protocols, failures in turn allows clients to become active only when the data
are ignored. The basic premise for that is that errors in of interest is being broadcast. Clients can remain in doze
fetching data as it becomes available in the channel
are deemed as tolerable, since the data will become avail-
able when it is rebroadcast. For real-time systems, however,
this assumption may imply missing critical deadlines. To
remedy these two problems, Bestavros [17] proposes a se-
ries of Broadcast Disk organizations that are suitable for
real-time environments and the usage of an Adaptive In-
formation Dispersal Algorithm (AIDA), which adds reli-
ability to the broadcast.
First, Bestavros presents three different organizations:
flat, rate monotonic and slotted rate monotonic which are
suitable for real-time environments. In flat organizations,
the simplest one of them, the server sends the union of the
data items needed by its client in a periodic fashion. With a
flat broadcast, there is the need of guaranteeing that the
worst-case latency will be tolerable even for those items
that have the strictest deadlines. In order to guarantee that, Fig. 3. Dispersal of two files.
enough bandwidth should be allocated. This organization,
is therefore, wasteful of bandwidth. To remedy this, the rate
monotonic organization broadcasts each item at a rate in-
versely proportional to its time constraint. So, each item
represents an independent disk “spinning” at a rate recip-
rocal to the timing constraint associated with the item. This
is obviously an extreme organization that tries to optimize
bandwidth at the expense of ease of broadcast program-
ming. The slotted rate monotonic organization achieves a
balance between the previous two organizations by coa-
lescing data items together so they could be put into the
same broadcast disk. This coalescence is achieved by parti-
tioning the data set into bins. Items in the same bin share
the same broadcast disk and therefore, have identical
worst-case latencies.
Fault-tolerance is achieved by Bestavros using AIDA, an
elaboration on the Information Dispersal Algorithm (IDA) Fig. 4. Broadcast program.
of Rabin [53]. Information Dispersal takes a file F and divides
112 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 11, NO. 1, JANUARY/FEBRUARY 1999

mode most of the time, thereby saving energy. In [36], [60], If the number of items exceeds the number of multicast
[37], the authors state the concept of Indexing on Air, i.e., addresses, the tuning time for accessing an item depends
transmitting an index along with the data, so clients can on the distribution of items among addresses. In [64], an
tune only during the times they need to. One issue that algorithm that optimally partitions the data is provided.
arises is how the index is multiplexed with the data to Broadcast schedules to minimize latency have been
make the latency and tuning time minimal. Sending an in- studied in [63]. The results obtained there can be combined
dex with every data broadcast increases the latency since with schemes to reduce tuning time as well. The main
clients that miss the index have to wait for the next broad- contributions of that study are a rule that characterizes
cast. Broadcasting the index m times during each data the minimal latency and three algorithms to schedule
broadcast one can minimize the probability of missing the broadcasts. The rule shows that the latency is minimized
index. This is called (1, m) indexing. The optimal value of m when the frequency of an item is proportional to the
along with the analysis of the technique is provided in [36]. square-root of its demand probability and inversely pro-
This technique can be further optimized by realizing that portional to the square-root of its size. The first scheduling
there is no need to replicate the entire index between data algorithm is an on-line algorithm that selects the item
2
segments. It is enough to send the portion of the index for which the expression (T − R(i)) pi/li is the largest. In the
relevant to the data segment that follows it. This is called expression, T is the current time, R(i) is the last time at
Distributed Indexing [36]. Distributed indexing results in which item i was transmitted most recently, pi is the de-
approximately the best latency and tuning time. It is also mand probability for item i and li the size of item i. This
possible to have secondary indexes on air, which are de- decision criteria can be computationally expensive. To de-
crease the computational demands, another on-line algo-
scribed in [60]. All these techniques, however, do not take
rithm is proposed in [63]. In this algorithm, items are di-
in consideration the access frequency of different data
vided in buckets which are kept as cyclical queues. The al-
items, or data skew in the construction of the index. Re-
gorithm chooses to broadcast the first item in the bucket for
ducing the average number of index probes can have a sig- 2
which the expression (T − R(Im)) qm/dm is the largest. In the
nificant impact in power consumption, so using the skew to
expression Im is the item at the front of the bucket m, while
guide the structure of the index is a good idea. This is the
qm and dm are the average demand probability and average
motivation of Chen et al. in [20], where imbalanced indexes
size for the items in bucket m. Simulations show that the
that favor most frequently accessed items are studied.
buckets algorithm performs close to optimal, specially as
The indexing techniques can be used to disseminate data the number of buckets is increased. Notice that the notion
using temporal addresses; that is, the client is given the exact of buckets can be compared to the notion of Broadcast
time when the relevant data item is going to be published. Disks [4]. The main differences, however are the facts that
Temporal matching is performed by the clock circuitry in the buckets algorithm is on-line and the algorithms in [4]
the client, consuming little energy. However, as a drawback, insists in placing the instances of items equally spaced in
this method requires a high degree of clock synchroniza- the disk. Finally an off-line algorithm is introduced in [63]
tion, forcing the network to be deterministic. This assump- by using an on-line algorithm to generate an a priori
tion is reasonable for exclusively used channels such as schedule of an specified size N.
Motorola’s Satellite Networks (Embarc) where only one Some recent applications of data dissemination include
server controls the channel. In general, however this is not information dissemination in the Internet [71], [18], [32],
easily achieved (consider for instance and Ethernet where and in private networks [28], [50], Advanced Traveler In-
the disseminating traffic can be mixed up with on-demand formation Systems [54], and dissemination using satellite
traffic). For these cases, the techniques have to be slightly networks [32], [22].
modified. One solution is to have clients tuning up ⑀ slots
ahead of time [60].
Alternatively, data dissemination can be achieved by
3 DATA CONSISTENCY
using multicast addresses [35]. The server sends data to a Consistency guarantees for data processed by mobile cli-
group of clients using the same address. In IP, for in- ents are an important area of research. These guarantees
stance, clients can use 32 bit addresses starting with 1111. provide the basis for any collaborative work and transac-
Equipped with an Ethernet card, a client can match the tion processing done with these systems.
predefined set of addresses and keep the CPU in doze In [61], the authors present a mechanism to provide
mode until the data’s multicast address matches the individual applications with a view of the database that
address specified by the client. Clients join multicast is consistent with their own actions. This is important
groups and filter the relevant data using the Ethernet card. since in their environment, clients can read or write data
Hashing can be efficiently used in combination with multi- from any one of the available servers and these servers
cast addresses. The client computes the hash value of the can contain inconsistent views of the database. For exam-
data item of interest and sets the interface card to listen ple, a mobile client of a distributed database could is-
to all broadcast values on the multicast address corre- sue a write at one server and later on read the item at
sponding to the hash value just computed. The hash func- a different server, obtaining an inconsistent value if the
tion can be given to the client upon registration. A ran- two servers have not been synchronized between the two
domized algorithm that minimizes the expected latency is operations. The paper introduces session guarantees that
described in [35]. alleviate this problem. A session is defined as a sequence of
BARBARÁ: MOBILE COMPUTING AND DATABASESA SURVEY 113

read and write operations performed during the execu- Using the demarcation protocol [10], one can ensure that
tion of an application by a client. The guarantees provided local executions are serializable. The protocol does this by
by their method are: exploiting the object organization and constraints. When a
• Read Your Writes: Any read operation in the session mobile client requires access to an object, it sends a request
must reflect the values established by previous writes to the database server invoking a split operation that selects
in that session. part of the object and establishes some consistency condi-
• Monotonic Reads: Successive reads reflect a nonde- tions. This part of the object is then only accessible by
creasing set of writes. transactions on the mobile client that requested it. The rest
• Writes Follow Reads: Writes are propagated after the of the object remains available for other clients. Examples of
reads on which they depend. objects that can be subjected to fragmentation are: aggre-
• Monotonic Writes: Writes are propagated after writes gate items (e.g., plane seats), sets and stacks. For instance,
that logically preceded them. sets can be split into a number of subsets and these subsets
recombined in an arbitrary order to reconstruct the original
These guarantees are enforced by having a session man- set. Each mobile host can specify a range of elements as the
ager that serializes read and write operations and by as- selection criteria, with the ranges being disjoint.
signing unique identifiers to each write. A server must be Gray et al. [29] propose a two-tier replication algorithm
willing to return information about the unique write identi- that allows mobile applications to propose tentative trans-
fier and the set of write identifiers for writes that are rele- actions at the mobile client. These tentative transactions are
vant to a given read. The session manager maintains a read- performed over the data replicas stored at the clients
set of identifiers for the writes relevant to session reads and while these clients are disconnected. When the mobile cli-
a write-set of identifiers for those writes performed in the ent reconnects, the tentative updates are reprocessed as
session. For instance, the Read Your Writes guarantee is base transactions in the node that stores the master copy of
enforced by following two steps. Whenever a write is ac- the data. The tentative transactions may fail when reproc-
cepted by a server, its identifier is added to the session’s essed (due to conflicts with other transactions). If that hap-
write-set. Before a read is processed in a server S at a time t, pens, the mobile client (and user) that originated the trans-
the session manager must check that the write-set is a sub- action is informed about the failure and why it happened.
set of the sequence of writes received by S before t. Processing the transaction in the master node may render
These session guarantees were developed to be deployed different from the tentative results. This may be acceptable
in the prototype implementation of the Bayou project, for some applications (checking account balances may
which aims to provide a collaboration framework between change because other transactions may have affected the
mobile computing users. balance) and unacceptable for others (an item is out of stock
Krishnakumar and Jain [42] discuss a method to perform and cannot be sold). This technique does a good job in
mobile sales transactions using site-transaction escrow solving a fundamental issue: standard ways of propagating
methods [10], [11], [41], [51]. Escrow methods divide the updates to replicas (eager replication [26], [25], lazy repli-
total number of available instance of an item among the cation [21], [56]) make deadlocks increase as the cube of the
number of sites in the system. A transaction can success- number of sites and as the fourth power of transaction size,
fully complete at a site if the number of instances it requires thereby rendering the systems unscalable. On the other
does not exceed the number available at that site. By using hand, the two-tier technique may result in an unacceptable
a reconfiguration protocol, such as the demarcation proto- number of failed transactions (after the clients reconnect)
col [10], sites can negotiate an increase of the allotted frag- and user dissatisfaction.
ment. In [42], a mobile client starts performing a transaction In [9], the concept of Certification Reports is introduced as
in one server, using escrow methods. If the client needs to a way of supporting transactions in a mobile environment.
move to another server during the life of the transaction The technique uses the broadcast channel to help mobile
(because it moved from one geographic area to another), clients do some of the work of verifying if the transactions
the only information needed at the new server is the opera- being run by them need to be aborted. Clients do that by
tion to be performed next. (Since the old server made its listening to Certification Reports that contain lists of items
decisions based on escrowed resources.) At the end of that are in the read and write set of active transactions that
transaction, however, a two-phase commit algorithm is have declared their intention to commit to the server during
needed to synchronize the two servers. This protocol is the previous broadcast period. If the client sees a conflict
therefore very suitable for clients that do a fair amount of between the reports read and writesets and its own read
mobility during the life of the transaction. and writesets, it aborts the transaction (this is inspired by
In [66], Walborn and Chrysanthis generalize the usage of optimistic concurrency control algorithms [43]). Otherwise,
escrow techniques by exploiting object semantics to facili- when the transaction is ready to commit, the client sends to
tate autonomous and disconnected operations in mobile the server the read and writesets along with the values to
database applications. The basic idea is to split large and be written in the database. The server does the final verifi-
complex objects into smaller fragments and cache a set of cation to check whether two transactions running on differ-
these fragments on a mobile host. By making these frag- ent clients collide with each other. The server then ac-
ments the unit of consistency and defining consistency con- knowledges accepted and rejected transactions using the
straints to be enforced in the fragments, one can have con- broadcast channel. One of the crucial advantages of this
current operations on a set of mobile computing clients. protocol is that it downloads most of the work of validating
114 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 11, NO. 1, JANUARY/FEBRUARY 1999

transactions to the clients. Moreover, transactions which to manage copies are also adequate to manage the direc-
ultimately would be rejected get an early abortion at the tory. However, a variation of the primary copy algo-
client site and can be resubmitted promptly (as opposed to rithm, called Primary by Row is proposed to manage direc-
the situation created by the technique in [29]. tories. In a Primary by Row algorithm, each core copy has
Lu and Satyanarayanan [47] propose a new transaction the ability to modify the portion of the directory that con-
model called isolation only transactions (IOT), with which the cerns the copy. For instance, in the example above, the core
system can provide strong consistency while not guaran- copy of the traveling document should be able to modify
teeing other traditional transaction properties, such as the row of the directory that points to the current location
atomicity and durability. The execution model uses opti- of the document.
mistic concurrency control [43]. As in the previous tech- A number of researchers are currently working in pro-
niques, a transaction is executed in the mobile client and viding replication in weakly connected systems, mainly
no partial execution is visible on the servers. When the using the Distributed Shared Memory paradigm ([45]).
transaction completes, it enters a committed or pending state. Among them, the work in [59] focuses in providing support
If the execution of the transaction does not contain any for very low-level objects of arbitrary sizes. The authors of
partitioned data access (i.e., the client maintains server con- [38] describe how to use Lazy Release Consistency [39] (which
nection for every item that has accessed), the transaction is is based on the fact that real consistency is not needed until
committed and its results made visible on the servers. Oth- a synchronization situation is reached), to develop a pro-
erwise, the transaction enters the pending state and waits to gramming model for large distributed systems with com-
be validated later. If the validation succeeds, its results are puting nodes that can be weakly connected.
reintegrated and committed in the servers. Otherwise, the
transaction is resolved (manually or automatically). In or-
der to do the validation, the technique applies Davidson’s
4 LOCATION DEPENDENT QUERYING
optimistic transaction model [23]. This method builds a The fact that clients in a mobile environment can change
precedence graph to represent interdependencies among locations enables the possibility of answering queries in a
transactions and detects cycles in the graph that are indica- way that is dependent on the current position of the client.
tive of transactions that do not satisfy global serializability. This is the subject of this section.
The last issue we address in this section is that of data Imielinski and Badrinath [33] presented the concept of
replication. The ability to replicate data in mobile environ- queries with location constraints, i.e., constraints which
ments is essential since disconnected units can continue to involve location of the mobile computing users. For in-
process objects that are stored locally. The authors in [12] stance the query “find me a doctor near my house” has a
point out that even though the number of copies in a sys- unary constraint on location. The query “find X, Y, and Z
tem can be large, the key issue is how to manages copies such they are in the same building and Y is between X and
that are updateable. This copies are also called the core copies Z” involves a ternary constraint (between) plus three unary
and their number is likely to be small for a simple reason: it constraints involving individual locations (in the same
becomes too expensive to update a large number of core building). The main problem becomes how to minimize the
copies. The number of read-only copies (or cached copies) communication cost to retrieve the necessary information to
can be large, but these copies may not be current. The core answer the query. Naive strategies result in too many mes-
set can be managed used standard quorum algorithms [26] sages and long latencies. On the other hand, building opti-
but very often the core set needs to be reconfigured. For mal plans can be shown to be NP-complete. The authors
instance, a person may be editing a document first from her suggest using greedy heuristics based on the ID3 [32] algo-
home computer, then from a laptop on a train ride and then rithm to solve the problem.
from the office computer. At each step a copy of the docu- Imielinski and Navas discuss a family of protocols that
ment is transferred from one computer to the next. At every integrate Global Positioning System (GPS)1 into IP to en-
step the copy in the computer being used currently is the able the creation of location dependent services [34]. Ex-
core set. In [67], Wang and Pâris describe a protocol to fa- amples of these services are: multicasting messages selec-
cilitate the reconfiguration of core sets. The protocol relies tively to specific geographical locations (such as a train sta-
in the introduction of nodes that act as referees. These nodes tion), providing services to clients within a certain geo-
are responsible for the updating and storage of the core set graphical range from the server and providing information
description. Referee nodes do not participate in data ac- for mobile users when that information depends on the
cesses and do not hold replicas of the objects. Their solely user’s location.
mission is to record changes in the composition of the core Genesis and Advanced Traveler Information Systems
set. Once a replica-holding node decides that it needs to (ATIS) [54], [55] are systems being developed to facilitate
change the core set, it sends the description of the new set travel. Travelers are mobile users equipped with personal
to the referees and if a majority of them accept the change, communication devices and portable computers. ATIS pro-
that description becomes the new core. Referees can be vides updated information on weather and road conditions,
highly replicated since they are only accessed when a new detours, construction zones, bus schedules, and parking.
core is elected. This information is available prior and during the trip. The
As it is pointed out in [12], the directory of objects is
also a replicated object that has to be managed just like 1. FPS is a radio-navigation satellite-based system developed and oper-
any other. The authors conclude that standard techniques ated by the United States Department of Defense. It provides users with
their three-dimensional geographical position, their velocity, and time.
BARBARÁ: MOBILE COMPUTING AND DATABASESA SURVEY 115

users can as a variety of queries, such as recommendations active badges [68], GPS, motion sensors and cameras, and
of the most favorable routes from a starting location to a human sources.
chosen destination. The term favorable may represent
minimum travel time, shortest distance, etc. Genesis is an
5 INTERFACES
ATIS operational test project developed in Minnesota as a
partnership between the University of Minnesota, the Min- The screen size of portable units has forced researchers to
nesota Department of Transportation, and private sector think about new interfaces for information retrieval.
companies. The Genesis system includes: data collection In [48], Massari et al. present a query processing inter-
face called Query By Icons (QBI), that addresses the issue
stations such as the Traffic Management Center (TMC) and
of screen size along with the limitations in memory and
Metropolitan Transportation Commission (MTC), the fixed-
battery power and the restricted communication band-
end ATIS database server which stores the data provided
width that are characteristic of wireless networks. The fea-
by the TMC and MTC along with the history of travel
tures of QBI are:
times, bus schedules and fares and other static information,
a wireless communication service provider and mobile per- • An iconic visual language that allows users to com-
sonal communication devices. Data is either disseminated pose queries by using a pointing device like a pen to
by the server as a result of a predetermined transaction or manipulate icons.
• A semantic data model that captures most of the as-
sent at the request of a mobile client. User queries include
pects of database structures.
requests for travel time, roads under construction in a given
• Metaquery tools that assist on the formulation of que-
area, shortest path for a trip, route evaluation and emer-
ries during periods of disconnection. This is done
gency service requests. Events, such as traffic accidents and
without having to access the actual data in the server.
road hazards are monitored and trigger data dissemination Data is accessed and transmitted back to the mobile
to travelers that need to be warned. Typical triggers warn client only when a complete query is answered.
travelers about congestion and alternative routing.
Mobisaic [65] is a World Wide Web information system The InfoPad is a project at the University of California,
Berkeley that aims to build a light weight, portable mul-
designed to serve mobile users. Mobisaic extends the
timedia notebook [44], [15]. The InfoPad main features are
WWW by allowing documents to refer and react to the cur-
portability, continuous network connectivity through a
rent location of clients. It does so by supporting two new
high-bandwidth radio link, pen input with handwriting
concepts: Dynamic Uniform Resource Locators (URLs) and
recognition, audio input with speech recognition and full-
Active Documents. A URL is dynamic if it contains at least a motion video playback with synchronized audio. One of
dynamic environment variable that gets substituted with the goals for the InfoPad is the development of a user in-
the environment context (e.g., location). For instance a user terface for application areas such as groupware and mo-
gets a hyperlink to a document that describes the user’s bile information browsing and retrieval [46]. This interface
current location in a building by placing a dynamic URL in addresses issues such as how to deal with the uncertainty
the corresponding WWW that contains the link. When a inherent in handwriting and speech recognition and how
user selects a dynamic URL in a document, the client the pen and voice can be used as substitutes for the mouse
browser is responsible for resolving all references to dy- and keyboard. Applications may use pen data from the
namic environment variables within the URL. After this InfoPad in three ways:
resolution, the browser has a standard URL which can be • First, regarding the pen as a mouse.
sent to the server to get the corresponding page. Active • Second, to get higher resolution data from the pen (an
documents are HyperText Markup Language (HTML) example of this can be a drawing program).
documents that enable the client to automatically react to • Third, using handwriting recognition to treat the pen
changes in a user’s mobile computing context. Authors as a keyboard.
write active documents just like they write standard HTML
Alternatively, applications may use a continuous-word,
documents, with the addition of a subscribe command that
speaker-independent speech recognizer to process audio
lists the dynamic environment variables that the client must data. To increase accuracy, the recognized can be custom-
subscribe when it loads the document. When the variables ized to individual speakers.
listed in the command change value, the client must take an Alonso and Mani [7], describe an implementation of a
action that reloads the current page (with new variables) or pen-based graphical database interface on a pen computer.
another page (marked by a new URL). Using a cellular phone, the interface connects to the remote
Spreitzer and Theimer [57] describe an architecture to databases and displays schema information. In response to
manage location information. In this architecture, per- gestures made by the user, the interface automatically gen-
sonal information is managed by User Agents, while a par- erates queries and retrieves results. For instance, the system
tially decentralized Location Query Service is used to fa- presents the user with the relations of a database. The user
cilitate location-based operations. There is a User Agent for can pick attributes of interest by simply tapping on them
each user. The agent collects and controls all personal in- with the pen. Furthermore, the user can ask to display all
formation regarding its user. Applications can only get per- possible join paths between attributes with another tap of
sonal information through a user’s agent. The sources of the pen or a P gesture. After that a join can be selected by
information collection for an agent include infrared-based tapping at the corresponding display and the resulting join
116 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 11, NO. 1, JANUARY/FEBRUARY 1999

table will be displayed along with the corresponding SQL [3] S. Acharya, M. Franklin, and S. Zdonik, “Balancing Push and Pull
for Data Broadcast,” Proc. ACM SIGMOD Int’l Conf. Management
query that represents it. The Universal Relation concept [40] of Data, Phoenix, Ariz., pp. 183-194, May 1997.
is used to aid in the automatic generation of queries based [4] S. Acharya, M. Franklin, S. Zdonik, and R. Alonso, “Disks on
on the attributes chosen by the user. Air,” Proc. ACM SIGMOD, 1994.
[5] S. Acharya, M. Franklin, S. Zdonik, and R. Alonso, “Broadcast
Disks: Data Management for Asymmetric Communication En-
6 CHALLENGES vironments,” Proc. ACM SIGMOD Int’l Conf. Management of Data,
San Jose, Calif., May 1995.
Although the volume of work in this area has been large [6] R. Alonso, D. Barbará, and H. Garcia-Molina, “Data Caching Is-
during the recent future, many challenges still remain. We sues in An Information Retrieval System,” ACM Trans. Database
point to some of them in this section: Systems, vol. 15, no. 3, Sept. 1990.
[7] R. Alonso and V.S. Mani, “A Pen-Based Database Interface for
• Prototyping: In the area of data dissemination, many Mobile Computers,” Proc. First IEEE Workshop on Mobile Comput-
theoretical studies, simulations, and focused imple- ing Systems and Applications, Santa Cruz, Calif., Sept. 1994.
[8] M. Ammar and J. Wong, “The Design of Teletex Broadcast
mentations have been tried. However, a full scale Cycles,” Performance Evaluation, vol. 5, 1985.
prototype that encompasses all of the main ideas is [9] D. Barbará, “Certification Reports: Supporting Transactions in
still missing. Wireless Systems,” Proc. 17th Int’l Conf. Distributed Computing Sys-
• Bandwidth utilization: A comprehensive study that ana- tems, Baltimore, May 1997.
[10] D. Barbará and H. Garcia-Molina, “The Demarcation Protocol: A
lyzes the best way to divide the bandwidth between all Technique for Mantianing Arithmetic Constraints in Distributed
the options (data broadcast, indexes, invalidation re- Database Systems,” Proc. Int’l Conf. Extending Data Base Technol-
ports, answers to queries and uplink capacity) is ogy, Vienna, 1992.
needed. The study in [3] is a good step towards un- [11] D. Barbará and H. Garcia-Molina, “The Demarcation Protocol: A
Technique for Maintaining Constraints in Distributed Database
derstanding the trade-offs, but more work is needed. Systems,” VLDB J., vol. 2, no. 3, 1994.
• Transactional properties: More work is needed in [12] D. Barbará and H. Garcia-Molina, “Replicated Data Management
studying real cases where transactional properties in Mobile Environments: Anything New Under the Sun?” Proc.
are needed in mobile applications. Moreover, study- IFIP WG10.3 Working Conf. Applications in Parallel and Distributed
Computing, Caracas, Venezuela, Apr. 1994.
ing which properties can be effectively relaxed (or [13] D. Barbará and T. Imielinski, “Sleepers and Workaholics: Caching
simply ignored) for real applications, as long as ef- Strategies in Mobile Environments,” Proc. SIGMOD, May 1994.
fective protocols to enforce the remaining properties [14] D. Barbará and T. Imielinski, “Sleepers and Workaholics: Caching
is an area that could have a very practical impact Strategies in Mobile Environments (Extended Version),” VLDB J.,
special the issue on the best of SIGMOD System-Oriented Papers,
in the field. vol. 4, no. 4, Oct. 1995.
• Optimization of location dependent query processing: Very [15] B. Barringer, T. Burd, F. Burghadt, A. Burstein, A. Chandrakasan,
little has been done in finding ways to obtain quick, R. Doering, S. Narayanaswamy, T. Pering, B. Richards, T. Truman,
and perhaps approximate, answers to these queries. J. Rabaey, and R. Brodersen, “Infopad: A System Design for Port-
able Multimedia Access,” Proc. Calgary Wireless Conf., July 1994.
• Data visualization: A key issue is to effectively use the [16] S. Baruah and A. Bestavros, “Pinwheel Scheduling for Fault-
scarce display space in mobile computers to present Tolerant Broadcast Disks in Real-Time Database Systems,” Proc.
answers to queries. 13th Int’l Conf. Data Eng., London, Mar. 1997.
[17] A. Bestavros, “AIDA-Based Real-Time Fault-Tolerant Broadcast
Disks,” Proc. Real-Time Technology and Applications Symp., Boston,
May 1996.
7 CONCLUSIONS [18] A. Bestavros and C. Cunha, “Server-Initiated Document Dis-
Mobile computing has proven a fertile area of work for semination for the WWW,” IEEE Data Eng. Bull., vol. 19, no. 3,
Sept. 1996.
researchers in the areas of database and data manage- [19] T.F. Bowen, G. Gopal, G. Herman, T. Hickey, K.C. Lee, W.H.
ment. The inherent limitations of mobile computing sys- Mansfield, J. Raitz, and A. Weinrib, “The DATACYCLE Architec-
tems present a challenge to the traditional problems of da- ture,” Comm. ACM, vol. 35, no. 12, pp. 71-81, Dec. 1992.
tabase management. [20] M. Chen, P.S. Yu, and K. Wu, “Indexed Sequential Data Broad-
casting in Wireless Mobile Computing,” Proc. 17th Int’l Conf. Dis-
As we can see, the amount of research in this area in the tributed Computing Systems, Baltimore, May 1997.
last few years have been staggering. However, some prob- [21] P. Chundi, D.J. Rosenkratz, and S.S. Ravi, “Deferred Updates and
lems remain open for research. There is a need for better Data Placement in Distributed Databases,” Proc. 12th Int’l Conf.
protocols in the area of data sharing and transaction man- Data Eng., New Orleans, 1996.
[22] S. Dao and B. Perry, “Information Dissemination in Hybrid
agement, better interfaces, clever algorithms that exploit Satellite/Terrestial Networks,” IEEE Data Eng. Bull., vol. 19,
locality to shape the answers to queries. Undoubtedly, we no. 3, Sept. 1996.
will continue to see a steady number of research contribu- [23] S.B. Davidson, “Optimism and Consistency in Partitioned Dis-
tions in the future. tributed Database Systems,” ACM Trans. Database Systems, vol. 9,
no. 3, pp. 456-481, Sept. 1984.
[24] M. Franklin, “Special Issue on Data Dissemination,” IEEE Techni-
cal Committee on Data Engineering, Sept. 1996.
REFERENCES [25] H. Garcia-Molina and D. Barbará, “How to Assign Votes in A
[1] S. Acharya, M. Franklin, and S. Zdonik, “Dissemination-Based Distributed System,” J. ACM, vol. 34, no. 4, pp. 841-860, Oct. 1985.
Data Delivery Using Broadcast Disks,” Personal Comm., vol. 2, no. [26] D.K. Gifford, “Weighted Voting for Replicated Data,” Proc. ACM
6, Dec. 1995. SIGOPS SOSP Conf., Pacific Grove, Calif., Dec. 1979.
[2] S. Acharya, M. Franklin, and S. Zdonik, “Prefetching from A [27] D.K. Gifford, “Polychannel Systems for Mass Digital Communi-
Broadcast Disk,” Proc. 12th Int’l Conf. Data Eng., New Orleans, cation,” Comm. ACM, vol. 33, no. 2, Feb. 1990.
1996. [28] D. Glance, “Multicast Support for Data Dissemination in Orbix-
Talk,” IEEE Data Eng. Bull., vol. 19, no. 3, Sept. 1996.
BARBARÁ: MOBILE COMPUTING AND DATABASESA SURVEY 117

[29] J. Gray, P. Helland, P. O’Neil, and D. Shasha, “The Dangers of [53] M.O. Rabin, “Efficient Dispersal of Information for Security, Load
Replication and a Solution,” Proc. ACM SIGMOD Int’l Conf. Man- Balancing, and Fault Tolerance,” J. ACM, vol. 36, no. 2, pp. 335-
agement of Data, Montreal, June 1996. 348, Apr. 1989.
[30] G.Y. Liu and G.Q. Maguire Jr., “A Mobility-Aware Dynamic Data- [54] S. Shekhar, A. Fetterer, and D. Lui, “Genesis: An Approach to
base Caching Scheme for Wireless Mobile Computing and Data Dissemination in Advanced Traveler Information Systems,”
Communications,” Distributed and Parallel Databases, special issue IEEE Data Eng. Bull., vol. 19, no. 3, Sept. 1996.
on databases and mobile computing, vol. 4, no. 3, pp. 271-288, [55] S. Shekhar and D. Lui, “Genesis and Advanced Traveler Informa-
July 1996. tion Systems,” Mobile Computing, T. Imielinski and H. Korth, eds.,
[31] G. Herman, G. Gopal, K. Lee, and A. Weinrib, “The Datacycle Kluwer, pp. 699-720, 1996.
Architecture for Very Large High Throughput Database Systems,” [56] J. Sidell, P.M. Aoki, S. Barr, A. Sah, C. Staelin, M. Stonebraker, and
Proc. ACM SIGMOD Conf., pp. 97-103, 1987. A. Yu, “Data Replication in Mariposa,” Proc. 12th Int’l Conf. Data
[32] Hughes Network Systems DirectPC, http://www.direcpc.com . Eng., New Orleans, 1996.
[33] T. Imielinski and B.R. Badrinath, “Querying in Highly Mobile and [57] M. Spreitzer and M. Theimer, “Providing Location Information
Distributed Environments,” Proc. 18th Int’l Conf. Very Large Data in A Ubiquitous Computing Environment,” Mobile Computing,”
Bases, Vancouver, B.C., Canada, Aug. 1992. T. Imielinski and H. Korth, eds., Kluwer, pp. 396-423, 1996.
[34] T. Imielinski and J.C. Navas, “GPS-Based Addressing and Rout- [58] K. Stathatos, N. Roussopoulos, and J.S. Baras, “Adaptive Data
ing, Computer Science Dept., Rutgers Univ., Technical Report No. Broadcasting Using Air-Cache,” Proc. First Int’l Workshop Satellite-
LCSR-TR-262, Mar. 1996. Based Information Services, Rye, N.Y., Nov. 1996.
[35] T. Imielinski and S. Viswanathan, “Adaptive Wireless Information [59] P.S. Knudsend M.O. Svendsen, and P.T. Koch, “Replication of
Systems,” Proc. SIG DataBase Systems Conf., Tokyo, Oct. 1994. Low-Level Objects in A Distributed Shared Memory,” Proc. Work-
[36] T. Imielinski, S. Viswanathan, and B. Badrinath, “Power Efficent shop Mobility and Replication, ECOOP ’95, Aug. 1995.
Filtering of Data on Air,” Proc. Fourth Int’l Conf. Extending Data- [60] T. Imielinski, S. Viswanathan, and B.R. Badrinath, “Energy Effi-
base Technology, Mar. 1994. cient Indexing on Air,” Proc. SIGMOD ‘94, May 1994.
[37] T. Imielinski, S. Viswanathan, and B. Badrinath, “Data on Air: [61] D. Terry, A.J. Demers, K. Petersen, M.J. Spreitzer, M.M. Theimer,
Organization and Access,” IEEE Trans. Data and Knowledge Eng., and B.B. Welch, “Session Guarantees for Weakly Consistent Repli-
1997, to appear. cated Data,” Proc. Conf. Parallel and Distributed Computing, Austin,
[38] C. Ionitoiu and B. Andersen, “Replicated Objects with Lazy Con- Texas, Oct. 1994.
sistency,” Proc. Workshop Mobility and Replication, ECOOP ’95, [62] D. Terry, D. Goldberg, D. Nichols, and B. Oki, “Continuous Que-
Aug. 1995. ries Over Append-Only Databases,” Proc. ACM SIGMOD Int’l
[39] P. Keleher, A.L. Cox, and W. Zwaenepoel, “Lazy Release Consis- Conf. Management of Data, San Diego, June 1992.
tency for Software DSM,” Proc. 19th Symp. Computer Architecture, [63] N.H. Vaidya and S. Hameed, “Data Broadcast Scheduling: On-
May 1992. Line and Off-Line Algorithms,” Dept. of Computer Science, Texas
[40] H.F. Korth, G.M. Kuper, J. Feigenbaum, A. Van Gelder, and J.D. A&M Univ., Techncial Report No. 96-017, July 1996.
Ullman, “System U: A Database System Based on the Universal [64] S. Viswanathan and T. Imielinski, “Wireless Data Access Using
Relation Assumption,” ACM Trans. Database Systems, vol. 9, no. 3, Multicasting,” Proc. Seventh Int’l Conf. Management of Data (CO-
pp. 331-347, Sept. 1984. MAD), Pune, India, Dec. 1995.
[41] N. Krishnakumar and A. Bernstein, “High Throughput Escrow [65] G.M. Voelker and B.N. Bershad, “Mobisaic: An Information Sys-
Algorithms for Replicated Databases,” Proc. 18th Int’l Conf. Very tem for A Mobile Wireless Computing Environment,” Mobile
Large Data Bases, Vancouver, B.C., Canada, Aug. 1992. Computing,” T. Imielinski and H. Korth, eds., Kluwer, pp. 375-395,
[42] R.M. Metcalfe and D.R. Boggs, “Ethernet: Distributed Packet 1996.
Switching for Local Computer Networks,” J. ACM, vol. 19, [66] G. Walborn and P.K. Chrysanthis, “Supporting Semantics-Based
July 1976. Transaction Processing in Mobile Database Applications,” Proc.
[43] N. Krishnakumar and R. Jain, “Mobility Support for Sales and 14th IEEE Symp. Reliable Distributed Systems, Germany, Sept. 1995.
Inventory Applications,” Mobile Computing, T. Imielinski and [67] Q.R. Wang and J.F. Pâris, “Managing Replicated Data Using
H. Korth, eds., Kluwer, pp. 571-594, 1995. Referees,” Proc. Workshop Mobility and Replication, ECOOP ’95,
[44] M.T. Le, S. Seshan, F. Burghardt, and J. Rabaey, “Software Archi- Aug. 1995.
tecture of the InfoPad System,” Proc. MOBIDATA Conf., Rutgers [68] R. Want, A. Hopper, V. Falcao, and J. Gibbons, “The Active
Univ., Oct. 1994. Badge Location System,” Trans. Information Systems, vol. 10,
[45] K. Li and P. Hudak, “Memory Coherence in Shared Virtual Mem- no. 1, Jan. 1992.
ory Systems,” Proc. Fifth ACM Symp. Principles of Distributed Com- [69] J. Wong, “Broadcast Delivery,” Proc. IEEE, vol. 76, no. 12, Dec. 1988.
puting, Aug. 1986. [70] K. Wu, P.S. Yu, and M. Chen, “Energy-Efficent Caching for Wire-
[46] A.C. Long, S. Narayanaswamy, A. Burstein, R. Han, K. Lutz, less Mobile Computing,” Proc. 12th Int’l Conf. Data Eng., New Or-
B. Richards, S. Sheng, R.W. Brodersen, and J. Rabaey, “A Proto- leans, pp. 336-343, Feb. 1996.
type User Interface for a Mobile Multimedia Terminal,” Proc. 1995 [71] T. Yan and H. Garcia-Molina, “SIFTA Tool for Wide-Area In-
Computer Human Interface Conf., May 1995. formation Dissemination,” Proc. Usenix Technical Conf., 1995.
[47] Q. Lu and M. Satyanarayanan, “Improving Data Consistency in
Mobile Computing Using Isolation-Only Transactions,” Proc. Fifth
IEEE HotOS Topics Workshop, Orcas Island, May 1995.
[48] A. Massari, S. Weissman, and P.K. Chrysanthis, “Supporting Mo- Daniel Barbará received his BSEE degree from
bile Database Access Through Query by Icons,” Distributed and Universidad Metropolitana in 1970 and the
Parallel Databases, special issue on databases and mobile comput- PhD degree in computer science from Princeton
ing, vol. 4, no. 3, pp. 47-68, July 1996. University in 1985. He is now an associate pro-
[49] B. Oki, M. Pfluegl, A. Siegel, and D. Skeen, “The Information fessor in the Department of Information and Soft-
BusArchitecture for Extensible Distributed Systems,” Proc. 14th ware Engineering at George Mason University
SOSP Conf., Asheville, N.C., Dec. 1993. in Virginia. His current areas of research in-
[50] B. Oki, M. Pfluegl, A. Siegel, and D. Skeen, “The Information terest include data warehousing, mobile sys-
BusAn Architecture for Extensible Distributed Systems,” Proc. tems, and data mining.
14th Symp. Operating Systems Principles, Asheville, N.C., Dec. 1993.
[51] P.E. O’Neil, “The Escrow Transactional Model,” ACM TODS, vol.
11, no. 4, pp. 405-430, Dec. 1986.
[52] R. Quinlan, “Induction of Decision Trees,” Machine Learning,
pp. 81-106, 1986.