A Caching Model for Real-Time Databases in Mobile

Ad-Hoc Networks
Yanhong Li and Le Gruenwald

School of Computer Science, University of Oklahoma

Norman, OK 73072, USA
{yanhong.li-1, ggruenwald}@ou.edu
http://www.cs.ou.edu/~database

Abstract. Although caching has been shown to be an efficient technique to improve the performance of database systems, it also introduces the overhead and
complexity in maintaining data consistency between the primary copies on
servers and the cached copies on clients. Little research has been performed for
data caching in the mobile ad-hoc network (MANET) environment where both
servers and clients are nomadic. In this paper, a caching model called
GMANET is designed to maintain both strong and weak cache consistency for
distributed real-time database transaction systems in group-based MANETs,
and at the same time, to incur as few update control messages as possible.
GMANET is compared with the existing caching models by means of simulation. The experiment results show that the GMANET has the best performance
in terms of percentage of transactions processed before their deadlines and is
compatible with other caching models in terms of mobile hosts energy consumption.

1 Introduction
With the advances in wireless networking technology and portable mobile devices, a
new computing architecture called mobile ad hoc wireless networks (MANETs) is
emerging. Applications in MANET are typically those that require the rapid deployment of mobile hosts and occur in a situation where a fixed infrastructure is not available. Example applications include military operations and disaster relief efforts.
Mobile hosts in MANETs are powered by short-lived batteries, communicate via
an unreliable wireless link, and move in various speeds. As a result, these mobile
hosts may experience severe network congestion, prolonged transaction execution, or
even frequent abortion of the transactions. These additional restrictions plus the deadline constraints imposed on time-critical applications call for a new power-aware and
communication-cost efficient caching technique for real-time MANET database system. Developing such a technique is the objective of our research.
Caching has been proven to be an essential technique for improving the performance of many computing environments, such as network file systems, wired distributed database systems, and web applications [3]. The purpose of caching is to bring
the data source as close to clients as possible, and thus, save a round-trip when the re*

This work was partially supported by the National Science Foundation grant No. IIS-0312746.

K.V. Andersen, J. Debenham, and R. Wagner (Eds.): DEXA 2005, LNCS 3588, pp. 186 196, 2005.
Springer-Verlag Berlin Heidelberg 2005

A Caching Model for Real-Time Databases in Mobile Ad-Hoc Networks

187

quested data are found in the local cache storage [1]. However, maintaining cache
consistency is a challenging problem. Cache consistency can be categorized into two
types: tight/strong cache consistency and loose/weak cache consistency [3, 14].
Strong cache consistency refers to the caching techniques that can always maintain
consistency between the cached data and the original ones. Weak cache consistency
refers to those that allow the data divergence between the cached data and the original
ones.
The rest of this paper is organized as follows. Section 2 reviews the current caching techniques in mobile database systems. In Section 3, GMANET, a caching model
for group-based MANETs, is proposed. Section 4 reports simulation results. Section 5
presents conclusions and future research work.

2 Literature Review
Several proposals have been made to solve the cache consistency problem in mobile
databases. The invalidation report technique in which the servers broadcast the invalidation reports to their clients periodically was proposed to maintain strong cache consistency [2,7]. But, it incurs the query latency [2,7] and tremendous communication
cost. The refresh time strategy aimed at maintaining weak cache consistency was proposed in [4]. However, it does not guarantee the freshness of the cached data so database transactions may access the dirty data.
Cooperative caching in MANET, which allows a client to access the cached data of
its neighbors, was proposed in [13]. The Time-To-Live (TTL) mechanism in [13] is
used to maintain the weak consistency level of the cache. Again, it is not applicable
for transactions that need accurate data. Another caching model, called MANET caching, was proposed in [10]. The refresh time strategy was adopted from [4] and modified to maintain the weak cache consistency. However, the cached data can be only
used for the read-only transactions that can tolerate out-dated data.
In reality, the mobile clients in many applications, such as battlefields, medical
emergencies, and fire-fighting operations, are organized in groups and their movements follow pre-defined patterns [6] instead of total randomness of mobility like the
mobility model used in [10]. Thus, a new caching model called GMANET for a
group-based MANET is proposed in this paper and also takes power consumption,
bandwidth, and real-time constraints into consideration.

3 The GMANET Caching Model

3.1 The GMANET Architecture
The GMANET architecture is illustrated in Fig. 1. Similar to the environment in the
MANET caching model in [10], the proposed group-based MANET also consists of
two representative devices, Large Mobile Hosts (LMHs) such as laptops, and Small
Mobile Hosts (SMHs) such as PDAs. The group-based MANET has a number of
groups. Each group logically has the following entities: group leader LMH (LMHg),
ordinary LMH and group member SMH. The LMHgs and LMHs have the whole database management system and SMHs has a caching and query processing module.

188

Y. Li and L. Gruenwald

The Location-Aided Routing protocol (LAR) [8] is assumed to carry out routing
packets from the sending MH to the receiving MH. The groups, group leaders, and
group members are defined by applications. If a group leader fails or needs to recharge its power, it will designate another LMH in its group as a deputy group leader
until it recovers and then assumes the group leader role again.
Hereafter in this paper, clients refer to those SMHs that initialize transactions and
send to servers, and servers refer to those LMHs that provide data service to other
network members. When LMHs request data from other servers, they themselves become clients.

Fig. 1. The GMANET architecture modified from [6]

3.2 Data Access and Update Model

The data model proposed in [10] is adapted in this paper. According to the update
characteristics, data are classified into the Periodically Updated data (PU) and Randomly Updated data (RU). The periodically updated data are the data that get updated
at fixed update intervals that are specified by applications. Obviously, the periodically
updated data are valid to use before their next updates. In battlefield applications, the
information about logistics of all battalions is an example of a periodically updated
data (PU). In contrast, the randomly updated data (RU) are updated anytime. An example of such data is the current number of refugees in the emergency and rescue operations. The purpose of this data update model is to support the different data freshness requirements of different transactions.
3.3 Transaction Type
In our real-time database system, transactions can be either firm or soft, which is defined by database applications. Firm transactions are aborted if they missed their
deadlines while soft transactions continue execution unless they missed their second
deadlines. Each transaction consists of a set of read and/or write operations. It has
been observed that most of the applications in mobile environments generate more
read operations than write operations [10]. All the write transactions are assumed to
be executed on the servers; while the read-only transactions can be executed by accessing the cached data items. In some applications, some read transactions might not
be as critical as the other read transactions with regard to data freshness. For example,
from a drivers viewpoint, the weather information and the traffic information measured at noon is not much different than those measured at 12:05 PM. But, the information about the location, emergency medical care, or accident investigation should

A Caching Model for Real-Time Databases in Mobile Ad-Hoc Networks

189

be as accurate as possible [9]. Therefore, the read-only transactions in [10] are further
divided into two types: 1) Up-to-Date (UD) type read-only transactions that accept
only fresh data (e.g. transactions querying data about locations of enemies) and 2) OD
(Out-Dated) type read-only transactions that accept slightly stale data (e.g. transactions requesting data about logistics of battalions).
3.4 The Caching Process in GMANET
3.4.1 The Cache Consistency
In the MANET caching model [10], the caches on the servers and clients are both
maintained at the weak consistency level by using the refresh time strategy, thus preventing all the UD type read transactions from using the cached copies. As a result,
The UD type read transactions have to be sent and distributed to the original servers,
lengthening the processing of these transactions. Therefore, the efficiency of the
MANET caching model depends largely on the application requirements. I
In our proposed GMANET caching model, both the strong and weak cache consistency levels will be maintained. The cache on clients will be maintained at the weak
consistency level by the refresh time strategy and the cache on group leaders will be
maintained at the strong consistency level by an asynchronous invalidation strategy.
GMANET with both cache types solves the above shortcoming of the MANET caching model. These two types of caching mechanism are discussed separately as follows.
3.4.1.1 Weak Cache Consistency on Clients
Clients are allowed to cache the previously accessed data items so that the subsequent
requests may be satisfied by the cached data and thus avoid sending them to servers.
The refresh time strategy in [4] is modified to keep the cached data consistent at the
weak consistency level on the client side (SMHs and ordinary LMHs) in the
GMANET caching model. Each cached data item is associated with a refresh time indicating how long this particular data item remains valid in the clients cache. In order
to calculate the refresh time, the update log containing the statistics about the update
pattern is maintained on each data server. The update log records the data id, the previous mean refresh time for this data item, and the latest update timestamp of all data
residing on the server. Before servers return the transaction results to clients, they estimate a refresh time for each data item in the transaction result and the estimated refresh time will be sent along with the transaction result back to clients.
The refresh time calculation is modified as follows to fit the GMANET caching

FRTi = Ti + d i Tcomm for randomly updated data, 2)

PRTi = Ti + Pi Tcomm for periodically updated data, 3) Tcomm = NoOfHops *
PackageSize / Bandwidth, where Ti is the current update timestamp of data item i,

model:

1)

d i is the mean update duration on data item i, Pi is the fixed update interval, Tcomm
is the transmission time between the sender and receiver, NoOfHops is the number of
hops between the sender and the receiver, PackageSize is the transmission amount,
and Bandwidth is the wireless bandwidth. The reason that the refresh time is sub-

190

Y. Li and L. Gruenwald

tracted by the communication cost is to reduce the effect of the communication time,
and thus, reduce the staleness degree of the cached data when transmitting from servers to clients.
Through the refresh time strategy the clients are not relying on the servers help to
validate their cached data items because each cached data item has already been attached with the refresh time specifying how long it is valid in the future. It means that
no additional communications between the clients and servers are needed. Unlike the
invalidation strategy where the clients have to be connected and tuned in to receive
the invalidation reports from servers periodically or asynchronously, this method allows clients to be free to move and disconnect (offline) and still be able to validate the
cached data items when the clients access their cached data. However, the refresh
time strategy cannot maintain strong cache consistency while the invalidation strategy
does.
3.4.1.2 Strong Cache Consistency on Group Leaders
Group leaders are allowed to cache the passing-by data on behalf of their clients. This
is because all the transactions initiated from clients are first sent to group leaders and,
therefore, group leaders can see all the network traffic within their registered clients.
Thus, LMHgs are selected as the locations where the cached data is maintained at the
strong consistency level and valid for access anytime. This will improve the processing of UD type read-only transactions.
The caches on group leaders are kept fresh by relying on the combination of invalidation and refresh time techniques. From the review on strong cache consistency
in Section 2, maintaining the cache at the strong consistency level is quite expensive
since it requires all the updates made on the servers be propagated to the cache holders immediately. In GMANET, of all the LMHs we assume only group leaders are allowed to maintain their cache at the strong consistency level for their clients, thus cutting down the total number of sites that need invalidation messages. As a result, the
communication overhead to maintain strong cache consistency is tolerable since the
number of group leaders is much smaller than the number of LMHs in GMANET.

Fig. 2. The invalidation messages from LMHs to LMHgs

The communication overhead to maintain strong cache consistency by invalidation

technique can be further reduced by making use of the refresh time for each cached
data item. As discussed in the previous section, each data server maintains an update
log, which is used to estimate refresh time for each accessed data item for its clients.
The data servers also use the refresh time to reduce the update or invalidation mes-

A Caching Model for Real-Time Databases in Mobile Ad-Hoc Networks

191

sage exchanges between them and the group leaders. As shown in Fig. 2, when data is
updated on an data server, it first looks up its update log to see whether the refresh
time of this updated data item is expired or not. If expired, then it is not necessary to
send the update message to group leaders; if not expired, then the data server must
send the update messages to each of the group leaders that have cached the updated
data item. For example, if 50% of the cached data are associated with the correct refresh time, then the update messages will be cut down by 50%, compared with the
cost of traditional invalidation methods to maintain strong cache consistency, where
every update invokes a communication for propagating the updated value to all the
cache holders.
It can also be seen that the tighter the estimated refresh time, the fewer update messages necessary to keep the cache consistent with the original servers. The tight estimation of the refresh time means that the refresh time tends to be small and expired
before the actual update, and thus the data servers do not need to send out update
messages, saving a lot of bandwidth and energy to transmit these control messages.
When a network partition occurs, the delayed update is assumed by the data servers. It means the data servers will wait to receive all acknowledgements of the group
leaders before the actual updates are committed. If some group leaders are in a different partition, the data servers will delay the updates until the refresh time expires.
3.4.2 Cache Data by Clients and by Group Leaders
When a transaction result is returned to a client from its group leader, the client will
check the transaction type since the client only caches the data requested by OD type
read transactions; the data items accessed by UD type read transactions have already
been cached by their group leaders. The reasons for this cache assignment are two
folds. First, it is to reduce the cache storage burden on the client side since clients are
much more limited in terms of memory and disk space. Second, it is to reduce the redundancy since it is not efficient if both clients and their group leaders cache the same
data. For each data item marked with the refresh time in the transaction result, if it is
already cached, the client updates its value and refresh time; if it is not cached and the
cache storage is not full, the client inserts it into an empty space; otherwise, the client
executes the replacement policy (Section 3.4.5) to select a slot for a newly arrived
data item.
The group leaders will cache only data that are accessed by UD read-only transactions. When the results of all the sub transactions are sent back from all the participating servers (LMHps), the group leaders have a decision to make as to what kind of
data they should cache. In order to reduce the overhead of the communication cost associated with the strong cache consistency policy, the group leaders only cache the
data accessed by UD type read-only transactions instead of all the transactions. The
accessed data by the OD type read-only transactions will be cached on the clients
side as discussed above. The cache on the group leaders will be used to satisfy all the
subsequent UD read-only transactions sent from their group members.
3.4.3 Cache Access by Clients
In GMANET, the cache on the client side is only kept at the weak consistency level,
and thus only OD type read transactions accepting slightly out-dated data can access it

192

Y. Li and L. Gruenwald

Fig. 3. The access of cached data items on the client side

(see Fig. 3); the cache on the group leaders is kept at the strong consistency level, and
thus, read transactions, either OD or UD type, can access the cached data on LMHgs.
All the write transactions have to be sent and processed at the group leaders LMHgs.
It is highly likely that the OD type read transactions are answered by using locally
outdated cached data because the weak consistency protocol is used on the clients
side. Even though the servers help estimate the refresh time based on the past update
statistics of the clients, the refresh time of each cached data item is only the best effort
estimation. Any updates happened on the servers may not reflect on the cached copies
of the clients in a timely fashion, and thus, it is inevitable for OD type read transactions to access the stale cache.
The access of the cached data items by clients is shown in Fig. 3. If an initiated
transaction is an OD type read transaction, the client checks its cache first and identifies all the cache misses and cache hits. It then sends the transaction requesting for
cache misses to the group leader. If it is the UD type read-only or write transaction,
the client sends it to its group leader for processing directly.
3.4.4 Cache Access by Group Leaders
When a transaction is initialized by an SMH and it is an OD read-only transaction, it
is first processed by the SMH to identify the locally qualified cached data items called
cache hit or cache miss otherwise. After processing, the SMH will package the cache
misses and send them to its group leader. Since the cached data on LMHgs are always
kept consistent with the original copies, all the read transactions, both UD and OD
types, can access the cached data.
The cache access by group leaders is shown in Fig. 4. When a new transaction arrives, and it is the read transaction, it will be first processed by using the cached data
maintained on the group leaders, and the cache hit data set, which is {i}, and cache
miss data set, which is {I} {i} {i} (note the {i} is the local cache hit data set on
the transaction initiator client), are identified. The group leader then distributes the
cache misses to all the participating servers (LMHps). All the write transactions are
distributed to the original servers for processing.

A Caching Model for Real-Time Databases in Mobile Ad-Hoc Networks

193

Fig. 4. The access of cached data on the group leader side

3.4.5 Cache Replacement

The cache replacement proposed in [10] is adapted in this research to make room for
newly arrived data items when the cache storage is full. Both the group leaders and
their group members use the same replacement policy. Initially, each mobile host
(server or client) will check whether its storage is full or not. If it is not full, it inserts
the newly arrived data into the empty entries in its cache. If it is full, it starts the replacement process by searching for expired data items in its cache. If an expired item
is found, it just replaces it; otherwise it searches for the data items with the lowest access frequency accessed by soft transactions. If it is not found, then it searches for the
data items with the lowest access frequency by firm transactions. The unique characteristic of this replacement policy is that it puts a higher caching priority on firm
transactions than soft transactions, and thus, favors firm transactions over soft transactions since the firm transactions must be aborted if they missed their deadlines.

4 Simulation Experiments
Four simulation models using AweSim software [11] are built to compare the
GMANET caching model, the MANET caching model, CHAN caching model and
the baseline model without caching module. The performance metrics are 1) the percentage of transactions missing their deadlines, indicating how many transactions
cannot be processed successfully within their transaction deadline requirements [10],
2) the total energy consumptions of all LMHs and all SMHs, 3) the average difference
in energy consumption between two LMHs indicating how balance the system is in
terms of energy consumption, and 4) the cache hit ratio, computed as the transactions
fulfilled by the servers or clients caches over the total transactions initialized in the
system. The parameters listed in Table 1 are the default parameter settings for the following simulation experiments.
4.1 Effect of Firm/Soft Ratio
The four models show different capacities of processing transactions before their
deadlines in Fig. 5. The NO caching model performs the worst, followed by the

194

Y. Li and L. Gruenwald
Table 1. The simulation parameters in the system
Value Range

Default

Firm/Soft Ratio

Parameter

0, 0.25, 0.5, 0.75, 1

0.5

OD/UD Ratio

0, 0.25, 0.5, 0.75, 1

0.75

0.2, 0.4, 0.6, 0.8

0.6

Number of LMHs

20

Number of SMHs

40

Simulation Area

1000x1000

Bandwidth
LMH Energy Dissipation
Rate in Active Mode
LMH Energy Dissipation
Rate in Doze Mode
SMH Energy Dissipation
Rate in Active Mode
LMH Energy Dissipation
Rate in Doze Mode

11 Mpbs

[12]

15.4 w

[5,12]

7.97w

[5,12]

2.178 w

[5,12]

1.4w

[5,12]

Cache Size

Reference

CHAN caching model and the MANET caching model; the best is the GMANET
caching model. With more firm transactions in the system, the performance gaps
among these four models become larger. It is obvious that the NO caching model is
the worst because it does not equip with the caching component like the other three
caching models. The better performance of the three caching models is contributed to
the caching mechanism so that some transactions can reduce their transaction paths by
going through shortcuts of the cached copies instead of the whole transaction path.
Also, the shortened transaction path means a reduced amount of transmission, thus
saving a lot of energy spent on transmission and computation. Among the three caching models, the GMANET has the best performance due to the fact that its caching
system allows not only OD type read transactions but also UD type read transaction to
access its cached data. Compared to GMANET, the MANET and CHAN Caching
models allow only OD type read-only transactions to access their cached data, and require all the UD type read-only transactions to be processed by the original servers.
F ig. 5 . E f f e c t o f F irm R a t io o n
% M is s e d D e a dline s
60

% Missed
Deadlines

50
40
30
20
10
0
0

0.25
0.5
0.75
1
F irm R a t io
NO Caching
M A NET
CHA N
GM A NET

A Caching Model for Real-Time Databases in Mobile Ad-Hoc Networks

195

4.2 Effect of OD Ratio

Fig. 6 shows that when the OD type transaction increases, the percentage of transactions missing deadlines decreases in both CHAN and MANET, but it remains more or
less the same in the GMANET caching model, and does not exhibit as much sensitivity to the changes of OD type transactions as the other two caching models do. The
above observed trend is expected. As all the caches in both the CHAN and MANET
caching models are maintained at the weak consistent level, only OD type transactions are allowed to access the cached data and all the UD type transactions will be
sent to the origin server to process. While the cache on the group leaders in
GMANET is maintained at the strong consistency level and the cache on the clients is
maintained at the weak consistency level, the OD type transactions can access the
cached data on the clients and on the group leaders, and the UD type transaction can
access the cached date on the group leaders. Thus, the performance gain in GMANET
is contributed to its double cache types in the system. The performance in terms of
percentage of transactions missing deadlines of GMANET is on average about 20%
better than MANET, 50% better than CHAN, and 100% better than NO Caching.
4.3 Effect of Cache Size

Fig. 6. Effect of Transaction OD/UD

Ratio on % Missed Deadlines

F ig. 7 . E f f e c t o f C a c he S ize o n
% M is s e d D e a dline s

40

40

% Missed
Deadlines

% Missed Deadlines

The cache size on the clients and servers is varied to show its effect on the performance of the four models. The NO caching model does not change with the cache size
in all the metrics studied because it has no caching mechanism. It can be seen in Fig.
7 that the percentage of the transactions missing deadlines decreases with the increase
of the cache size in all the three caching models. With the increase of the cache size
more data can be cached for the subsequent transactions, and thus, the probability of
satisfying the transactions with the local cached data increases. Among the three caching models, the GMANET technique performs the best, followed by MANET and
CHAN. The same analysis as that in the previous experiments holds true for the performance differences among the three caching models.

20

20

0
0

0.25

NO Caching
CHAN

0.5
OD Ratio

0.75
MANET
GMANET

0.2

0.4
0.6
C a c he S ize

NO Caching
CHA N

M A NET
GM ANET

0.8

196

Y. Li and L. Gruenwald

5 Conclusions
Designing a caching technique in group-based MANETs is meaningful since, in practice, with most applications such as battlefields, medical emergencies, and firefighting, there are several logical units involved and their movements follow some
pre-defined patterns instead of total randomness. A new caching model called
GMANET has been proposed for group-based MANETs in this research, and the preliminary simulation results show that it performs the best in terms of its ability to
complete transactions before their deadlines. However, a hand-off mechanism is
needed to handle the scenario when the group leaders run out of energy and new
group leaders are designated as their successors.

References
1. Baentsch, M., Baum, L., Molter, G., Rothkugel, S., and Sturm, P.: Enhancing the Web's infrastructure: from caching to replication. IEEE Internet Computing. 1 (1997) 18-27
2. Cao, G.: On Improving the Performance of Cache Invalidation in Mobile Environments.
Mobile Networks and Applications, 7(2002) 291-303
3. Cao, L.Y., and OZSU, M.T.: Evaluation of Strong Consistency Web Caching Techniques.
World Wide Web: Internet and Web Information Systems, 5 (2002) 95-123
4. Chan, B-Y., Si, A. and Leong, H-V.: A Framework for Cache management for Mobile Databases: Design and Evaluation. Distributed and Parallel Databases, 10 (2001) 23-57
5. Feeney, L. M., Nilsson, M.: Investigating the Energy Consumption of a Wireless Network
Interface in an Ad Hoc Networking Environment. In Proc. of IEEE INFOCOM. Anchorage, Alaska, (April 2001) 1548-1557
6. Hong, X., Gerla, M., Pei, G., Chiang, C-C.: A Group Mobility Model for Ad Hoc Wireless
Networks. In Proc. of the 2nd ACM intl workshop on Modeling, analysis and simulation of
wireless and mobile systems. Seattle, Washington (August 1999) 53-60
7. Jing, J., Elmagarmid, A., Helal, A. and Alonso, R.: Bit-sequences: An Adaptive cache invalidation method in mobile client/server environments. Mobile Networks and Applications, 2 (1997) 115-127
8. Ko, Y-B., Vaidya, N. H.: Location-Aided Routing (LAR) in mobile ad hoc networks. Wireless Networks, 6 (2000) 307-321
9. Lam, K-y., Chan, E., Yuen, Joe C-H.: Broadcast Strategies to Maintain Cached Data for
Mobile Computing System. In Proc. of the Workshops on Data Warehousing and Data
Mining: Advances in Database Technologies. Springer-Verlag, London, UK(1998) 193204
10. Lau, C.: Handling Mobile Host Disconnection, Data Caching, and Data Replication in
Managing Real-Time Transactions for Mobile Ad-Hoc Network (MANET) Databases.
Master Thesis, University of Oklahoma, Norman, OK (2002)
11. Prisker, A. Alan B., OReilly, Jean J.: Simulation with Visual SLAM and AweSim. Wiley
System Publishing Corporation, West Lafayette, Indiana (1999)
12. Shih, E., Bahl, P., Sinclair, M.J.: Wake on Wireless: An Event Driven Energy Saving Strategy for Battery Operated Devices. In Proc. of the Eighth Annual Intl Conf. on Mobile
Computing and Networking. Atlanta, Georgia (September 2002) 160-171
13. Yin, L. and Cao, G.: Supporting Cooperative Caching in Ad Hoc Networks. In Proc. of the
5th ACM intl workshop on Wireless Mobile Multimedia. Atlanta, Georgia(Sep.2002) 56-63
14. Yu, H., Breslau, L., Shenker, S.: A Scalable Web Cache Consistency Architecture. In ACM
SIGCOMM Computer Comm. Rev., Proc. of the conf. on Applications, technologies, architectures, and protocols for computer communication. Cambridge, MA (1999) 163-174