PReCinCt: A Scheme for Cooperative Caching in Mobile Peer-to-Peer Systems ∗

Huaping Shen, Mary Suchitha Joseph, Mohan Kumar, and Sajal K. Das
Center for Research in Wireless Mobility and Networking (CReWMaN)
Department of Computer Science and Engineering, University of Texas at Arlington
Arlington, TX 76019, USA
{hpshen, mjoseph, kumar, das}@cse.uta.edu

Abstract wireless exchanges of data [10]. Typically, P2P systems en-

Mobile Peer-to-Peer (MP2P) systems consist of mo-
bile peers that collaborate with each other to complete
application problems. Information sharing in such en-
vironments is a challenging problem due to the fun-
damental limitations of battery power, wireless band-
width, and users’ frequent mobility. We proposed a novel
scheme, called Proximity Regions for Caching in Coopera-
tive MP2P Networks (PReCinCt) to efficiently support scal-
able data retrieval in large-scale MP2P networks. In the
PReCinCt scheme, the network topology is divided into ge-
ographical regions where each region is responsible for
a set of keys representing the data. In this paper, we ex-
tend the PReCinCt scheme to facilitate consistent cooper-
ative caching in MP2P systems. The caching scheme con-
siders data popularity, data size and region-distance dur-
ing replacement to optimize cache content of peers.
PReCinCt employs a hybrid push/pull mechanism to main-
tain data consistency among replicas in the network. Sim-
ulation results show the cost of consistency maintenance
in terms of latency and energy consumption is signifi-
cantly improved in the PReCinCt scheme.
1. Introduction and Background
In Peer-to-Peer (P2P) systems, peers with equiva-
lent functionality, communicate and share resources. With
the rapid advancements in mobile wireless communica-
tion technology, P2P computing has also been introduced
into mobile and wireless networks, thus leading to mo-
bile peer-to-peer (MP2P) networks. MP2P systems con-
sist of peers that are user applications executing on mobile
devices. Mobile peers interact during (brief) physical en-
counters in the real world, thereby engaging in short-haul
∗ This work was supported by the National Science Foundation under
Grant No. 0129682 and the Texas Advanced Research Program under
Grant Number: 14-771032.
able direct real-time sharing of services and informa-
tion among distributed peers. In contrast to P2P systems
on wired networks that comprise static peers, MP2P sys-
tems are subjected to the limitations of battery power, wire-
less bandwidth, and highly dynamic network topology.
Thus MP2P systems give rise to new challenges for re-
search on routing, resource discovery, data retrieval, data
consistency maintenance, security and privacy manage-
ment.
In existing unstructured P2P networks [17], flooding is
the most popular data retrieval mechanism. Flooding en-
tails message processing at every node which is expensive
in terms of bandwidth, battery power and computational re-
sources. In [12], the expanding ring scheme is proposed to
reduce the cost of each data retrieval. A node starts a flood
for a request with a small Time-to-Live (TTL) and contin-
uously increases the TTL until the data is found. Due to
the resource constraints (e.g., battery and wireless band-
width), both flooding and expanding ring schemes are not
expected to adapt well in large scale MP2P networks. In the
structured P2P networks [15], distributed hash table (DHT)
is used to deploy and retrieve data in controlled network
topologies. In DHT based data retrieval schemes, given a
key, the query will be routed to a specific node (location)
that is responsible for storing the value associated with the
key. However, due to the nodes’ frequent mobility, the data
retrieval schemes of structured P2P networks are also not
expected to perform well in MP2P networks.
In this paper, we investigate cooperative caching and
consistency issues in a novel scheme, called Proxim-
ity Regions for Caching in Cooperative MP2P Networks
(PReCinCt) [11]. In this scheme, the entire network topol-
ogy is divided into geographical regions, each being
responsible for a set of keys representing the data of in-
terest. A hash function h(ki ) = Lj is used at each peer to
map a key ki to a region (Rj )’s location Lj . We call the re-
gion Rj as the home region of key ki . When a peer needs a
data item represented by a key ki , the peer obtains the lo-

Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
 cation information of home region Rj by using the hash
function, and then a request message is sent to the home re-
gion Rj by using a geographic-aided routing protocol,
such as GPSR [8]. After reaching the home region, lo-
calized flooding is used to locate the peer holding the
requested data. By routing to regions rather than to spe-
cific points, PReCinCt requires only approximate location
information of each region, thus making it robust to er-
rors in location measurement and frequent mobility of
peers.
Caching is a key technique for improving the data re-
trieval performance of mobile devices [14]. Cooperative
caching is a mechanism in which peers can access data
items from the cache in their neighboring peers. In the lit-
erature, cooperative caching schemes have been studied to
improve the performance of web proxies in wired networks
[5][13]. In order to further save bandwidth for each data re-
trieval, PReCinCt incorporates a novel cooperative caching
scheme that caches relevant data among a set of peers in
a region. In contrast to existing caching schemes, our co-
operative caching scheme determines cache replacement by
considering not only each peer’s own access frequency but
also the importance of data items to other peers in the same
region. The proposed caching strategy, Gready-Dual Least
Distance (GD-LD), uses a utility function to evaluate the
importance of each data item based on a combination of
three factors: the popularity of the item in the region, size
of the item, and the region distance between the request-
ing and responding peers. In this paper, the words peer and
node will be used interchangeably.
Most existing P2P systems are specifically designed to
share static content such as music and video files and thus
assume that data is static and updates occur very infre-
quently [4]. To handle dynamic data, the existing techniques
must advance from a primarily read-only system to one in
which data can be read as well as updated frequently. Han-
dling dynamic data necessitates effective consistency main-
tenance schemes to ensure that all replicas of a data item are
consistent with each other. These observations have moti-
vated us to consider cache consistency issues in MP2P sys-
tems.
In this paper, we propose a hybrid push/pull algorithm
called Push with Adaptive Pull that uses minimal message
overheads to maintain data consistency. We investigate and
compare the proposed scheme with two existing cache con-
sistency algorithms: Plain-Push [3], which is a push based
scheme initiated by the peer making the update; Pull-Every-
time [7] initiated by the peer requesting for the data. The
proposed push with adaptive pull algorithm enhance the
original PReCinCt
In Plain-Push, the new update for a data item is pushed
by the initiator to the entire MP2P network using flooding.
When a peer receives an invalidation message, it checks its
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
cache and invalidates the relevant data. The advantage of
the push scheme is its simplicity and stateless nature. How-
ever this scheme substantially increases the control message
overhead due to subsequent broadcasts of the invalidation
messages and entails message processing at every peer. This
is undesirable in mobile environments where bandwidth is
scarce and peers have very limited battery power. In addi-
tion if a peer is unreachable, it will not receive the invalida-
tions thus introducing the risk of violating consistency guar-
antees.
In the Pull-Every-time scheme individual peers are re-
sponsible for maintaining consistency of their cached items.
In this scheme, each time a peer P requests for a data item
d and there is a local cached copy, the peer polls the owner
of the data to validate its copy of d. The advantage of this
scheme is its simplicity and strong consistency guarantees.
The disadvantage is that it generates a large amount of net-
work traffic thus incurring a considerable amount of band-
width. Also, the latency increases as the response time in-
curs an additional round-trip delay to poll the owner of the
data item.
The Push with Adaptive Pull, combines the positive fea-
tures of the Plain-Push and Pull-Every-time schemes in or-
der to address the problems of the original schemes. Dur-
ing the push phase of our proposed scheme, the peer up-
dating a data item sends the update to the home region for
that data item. During the pull phase, peers poll the home
regions to check if their cache contents are obsolete. We
propose an adaptive polling scheme that determines the fre-
quency at which peers should poll the home regions. The
home region assigns a Time-to-Refresh (TTR) value for its
data items. This TTR value is an estimation of the duration
of time during which the data item is less likely to change.
Requests for a data item before the expiration of the TTR
are served by the peer’s local cache. Upon expiration of the
TTR, peers poll the home region to check for the validity of
the data item.
The rest of the paper is organized as follows: Section
2 presents an overview of the PReCinCt scheme. This
section includes a discussion on region management and
fault-tolerance and data replication. Section 3 discusses the
caching strategy employed in PReCinCt and Section 4 dis-
cusses data consistency issue in PReCinCt scheme. Section
5 presents analytical performance evaluation of PReCinCt
scheme. Section 6 presents detailed simulation results and
finally Section 7 concludes the paper and discusses our fu-
ture work.
2. Description of PReCinCt Scheme
This section describes our PReCinCt scheme for data
retrieval in MP2P networks. In this scheme, the network
topology is divided into geographical regions, each respon-
 sible for a set of keys of data. The keys for the data are dis-
tributed among the regions, such that each key ki maps to a
region Rj . This key-set to region mappings are known to all
peers by using a geographic hash function. Every data item
is associated with a home region, where the data item is ini-
tially stored.
2.1. Region Management
In PReCinCt, the whole network topology is divided into
multiple geographical regions. Each region is represented
by the location information of its center point and all ver-
tices in perimeter. Each peer uses a region table to keep the
location information of all regions in the whole network.
When a peer joins the network for the first time, the peer can
retrieve the region table from its neighboring peers. The size
and shape of the regions can be changed by updating the re-
gion table of all peers in the network. At the same time,
each key in the network also needs to be relocated accord-
ing to the region table changes.
There are four operations that can be applied to modify
regions - Add, Delete, Merge, and Separate. In Add opera-
tion, a new entry which includes the location information of
new region is added into the region table to indicate the ex-
pansion of the whole network topology. When a region is
no longer in the network, the Delete operation removes the
entry of corresponding region from region table. Merge re-
places the entries of two existing regions in the region ta-
ble with the location information entry of the new region,
indicating the merging of two neighboring regions. Sepa-
rate is used to divide one region into two new regions. All
the four operations are executed by updating the region lo-
cation information in the region table. After each execution,
the peer needs to disseminate the update to all other peers in
the whole network to guarantee the consistency of region ta-
bles of all peers.
2.2. Data Search Process
When a peer needs a data item, it first floods the request
in the region in which it currently resides to determine if
any of its neighboring peers have a local cached copy of the
item. If so, it is a local hit, the request succeeds and the re-
sponse is sent back to the requesting peer. If not, the re-
questing peer uses the geographic hash function to get a lo-
cation according to the query. Home region is determined
by searching the region whose center location is closest to
the location from the region table. Then the requesting peer
generates a request message containing the following three
fields: i) the identity of the peer making the request, ii) the
location of home (destination) region for the requested item,
and iii) the key of data item being requested.
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
Nodes routing the requested message towards the desti-
nation region check the destination region location in the
header, to determine whether they are within that region.
The first node inside the destination region receiving the re-
quest message, identified as the point of broadcast, floods
the message within the region to locate the peer holding
the data item. Each peer in the home region processes the
request message to determine if it has the requested data.
Peers located outside the home region drop the request
message without further processing. When the data is lo-
cated, the response is sent back to the original request-
ing peer and the search process terminates. When the re-
questing peer receives the response, the cache manager de-
cides if this data item should be cached. (Cache manage-
ment will be discussed in the next section). Thus using the
PReCinCt scheme, flooding is limited to within a region
leading to savings in network bandwidth and energy con-
sumption at the nodes. The detailed algorithm of search pro-
cess of PReCinCt is given in Figure 1.
2.3. Peer Mobility Handling
In PReCinCt, mobility of peers is classified into cate-
gories: intra-region and inter-region. In the intra-region mo-
bility, a peer moves in the same region, thus the overheads
are minimal. In the inter-region mobility, on the other hand,
a peer moves out of its initial region to a neighboring re-
gion. Peers check their positions periodically to detect a
inter-region mobility. If a inter-region mobility occurs, the
peer has to distribute its keys to other peers in its origi-
nal region. To reduce the overhead of inter-region mobil-
ity, the peer moving out of the region sends its keys to the
other peers such that: 1) have low mobility rates; 2) are lo-
cated near the center of the region; and 3) have cache space
to store this data. Peers with low mobility and those located
near the center of the region have a low probability of leav-
ing the region in the near future.
2.4. Fault Tolerance and Replication
There is one major obstacle that all peer-to-peer sys-
tems must overcome - that of node disconnecting or sud-
den crash. Due to the high peer mobility and low reliability
of wireless links, the problem is more severe in the MP2P
systems. We must therefore make fault-tolerance a priority.
A general solution to fault tolerance problem is is very
hard to construct. We make three assumptions based on our
target application and expected users’ behavior in MP2P
systems. i) Node failures are independent: Since nodes are
owned by different users and nodes use peer-to-peer wire-
less links, we assume that a set of nodes with adjacent geo-
graphic locations are highly unlikely to have dependent fail-
ures. ii) Most users quit the network gracefully: If a mov-
 ing node retains some keys, such keys will be transferred to
Notations:
other nodes in the same region prior to disconnection of the d: requested data item;
node from the MP2P system. iii) Messages will be routed Preq : peer requesting d;
to the correct node: Despite node mobility and inhomoge- Rreq : region in which Preq resides;
Presp : peer which responds with data d ;
neous node distribution in MP2P networks, it is reasonable Rresp : region in which Presp resides;
to assume that messages eventually reach the correct node. U (i): Utility value of data item i;
Most underlying ad-hoc routing protocols provide lim-
Procedure Search(d)
ited fault-tolerance to make the routing resilient to user mo- Begin
bility and node failures, but data items in MP2P systems still if (d is cached in Preq )
need to be replicated to improve their availability. Further- Update utility value of d in Preq ;
return d;
more, the replicas must be kept consistent with the original else
after each update. Since PReCinCt keeps keys at their home Broadcast the request for d in Rreq
regions, we define the home region failure as the failure that for each (peer Pi in Rreq )
if (d cached in some peer Presp )
there is no copy of a key in its home region. Home region Update utility value of d in Presp ;
failure happens when there is no peer in a region, a peer Send response to Preq
did not send back the key to the home region after a inter- else
Obtain the home region Rresp for data d;
region mobility, or sudden death occurs to the peer hold- Route the request for d to Rresp
ing the key. Based on the above three assumptions, we de- Broadcast the request in Rresp
sign a lightweight data replication mechanism in PReCinCt if ( d cached in some peer Presp )
Update utility value of d in Presp ;
to tolerate network, nodes and home region failures. The al- Send response to Preq
gorithm tries to maintain at least one replica of each key in End
a region other than the home region under dynamic circum-
Procedure CacheAdmissionControl(d)
stances to improve the availability of each data item. The Begin
algorithm can be easily extended to to multiple replicas for if (Rresp = Rreq )
coping with higher failure frequencies. Multiple copies re- Calculate the utility of data d;
if (cache space available for d)
sult in increased bandwidth usage for replica messages and cache d;
more cache space. else
call CacheReplacementPolicy(d);
As we mentioned in the Data Search Process section, else
home region of each key is determined by selecting the do not cache d;
region whose center location is closest to the key’s hash End
value. Given a data item with key K whose hashed value Procedure CacheReplacementPolicy(d)
is location L, and the home region Rh is the region whose Begin
center location (Lh ) is closest to L. We will similarly se- L = min(U (q)), ∀q in the cache
Evict q which satisfies U (q) = L and cache d;
lect the next closest region Rr whose center location is Lr U (d) = L + U (d);
as a replica region of K, i.e., ∀Ri : dist(L − Lh ) ≤ End
dist(L − Lr ) ≤ dist(L − Li ). This means that a request
message for key K will always be routed to the region Rh . Figure 1. Algorithms of the PReCinCt
If there exists network failure, node failure or region failure Scheme
of Rh , this request message will instead be rerouted to the
replica region Rr . After reaching the Rr , localized flood-
ing is used to locate the requested data item. In order to tency of subsequent retrievals. The dynamic cache is opti-
maintain the consistency between original key and replica, mally managed by a greedy cache mechanism. In this sec-
update message needs to be sent from home region to the tion we present the caching scheme employed in PReCinCt
replica region if the home region receives a data update for that includes a cummulative cache, a cache admission con-
key K. We will discuss the details in the next section. trol and a cache replacement policy.

3. Cooperative Caching in PReCinCt 3.1. Cumulative Cache

The cache space of each peer is divided into two parts:
static and dynamic. The static space contains the values of
keys that belong to the region where the peer currently re-
sides. The dynamic space, on the other hand, contains data
items that are placed opportunistically at a peer to reduce la-
Caching data items in the local caches of peers, helps re-
duce latency and increase accessibility. When a peer Pi re-
quests for data item d, it first attempts to obtain the data lo-
cally from its own region by broadcasting the request to its
regional members. A local hit occurs if the request can be

Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
 satisfied within the same region. If not, the request is sent to
the home region of that data. If a peer along the path to the
home region has the requested data item d, then it serves the
request without forwarding it further towards the home re-
gion. Otherwise, the request is forwarded to the home re-
gion of item d. Since the local caches of the peers virtually
form a cumulative cache, decisions regarding the caching of
a data item and its eviction from the cache depend not only
on the peer itself but also on the neighboring peers. There-
fore, PReCinCt includes a cache admission control policy
and a cache replacement algorithm.
3.2. Cache Admission Control
When a peer Pi receives the response for the requested
data item, a cache admission control is triggered at Pi to de-
cide whether the data item should be cached. If the origin
of the data resides in the same region as the requesting peer
Pi , then the item is not cached; otherwise it is cached at the
peer Pi . Peers cooperatively cache data and thus it is unnec-
essary to replicate data in the same region, as they can be
obtained locally for subsequent requests.
3.3. Cache Replacement Policy
We have developed a greedy replacement policy which
considers three factors while selecting a victim: i) the ac-
cess count of the data items that reflects the popularity of
a data item in the region; ii) the size of the data items; and
iii) the region-distance which is the distance between the re-
gions of the requesting and the responding peers. This algo-
rithm incorporates the region-distance as an important pa-
rameter in selecting a victim for replacement. The greater
the region-distance, the greater is the utility of the data.
This is because caching data items which are further away,
saves bandwidth and reduces latency for subsequent re-
quests. Thus we call our replacement algorithm Greedy-
Dual Least-Distance (GD-LD).
The cache replacement policy employs a utility function
to assign a utility value for each data item, based on the
three factors mentioned above. Since the nodes in the re-
gion cooperatively cache data, the utility value reflects the
importance of the data item to the entire region. The util-
ity value is calculated using the following expression:
1 End
U tility = wr × aci + wd × reg dst + ws × (1)
szi
where aci is number of times the item di has been accessed
in the region; reg dst is the distance between the request-
ing region and the home region for the data; szi is size of the
item di , and wr , wd , ws are the corresponding weight fac-
tors. The utility value of the data item is updated when there
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
is a hit. This value is used by the greedy replacement algo-
rithm at the peer to find an optimal replacement and thereby
manage the cache efficiently.
4. Data Consistency in PreCinCt
In this section we discuss our proposed Push with Adap-
tive Pull scheme to maintain data consistency among repli-
cas in the MP2P network. This scheme combines the push-
based invalidation initiated by the peer updating a data item
with the pull based invalidation initiated by the individual
peers that maintain the cache. During the push phase the
peer updating a data item sends the update to the home re-
gion for that data item. During the pull phase, peers poll
the home regions to check if their cache contents are obso-
lete. The proposed an adaptive polling scheme determines
the frequency at which peers should poll the home regions.
Push Phase: When a peer Pupdate initializes an update
request for a data item d at time t, Pupdate updates d and
pushes the update message to the home region Rh of d as
well as replica region Rr of d. Within the home and replica
regions flooding is used to locate the peer, Pi that has the
data item d. Pi updates its copy of data item d. Thus by
pushing the update to only the home and replica regions in-
stead of to the entire network, our scheme saves on band-
width of the wireless network and incurs less message over-
head to maintain data consistency. After receiving the up-
date message, Pi updates the Time-to-Refresh (TTR) value
for d. The TTR values are dynamically maintained to re-
flect the update rates for the data items. The algorithm of
the Push phase is given in Figure 2.
Notations:
d: updated data item;
t:time at which the update is initiated;
Pupdate : peer initiating d;
Rh : home region for data d;
Rr : replica region for data d;
M (d,t): Update message for data d which is updated at time t;
Procedure Push
Begin
When Pupdate initiates an update message M (d, t)
Obtain the home region Rh and replica region Rr for d
Route the message M (d, t) to regions Rh and Rr
Within Rd or Rr locate the peer Pi which has d
Update d and store the update time t in Pi
Pi calculates TTR of data item d
Figure 2. Push Phase
Pull Phase: The pull phase of the algorithm is initiated
by the individual peers interested in checking for the valid-
ity of their cached items. During this phase peers poll the
 home regions to determine the freshness of data items in
their cache. We propose an adaptive pull mechanism.
Adaptive Pull: The notion of adaptive polling has been
used in the context of web cache consistency [7] and we
adopt a similar idea here. In our proposed adaptive polling
mechanism, peers vary the frequency with which they poll
the home region for a data item, depending on the update
rate for that data item. In this scheme the polling frequency
is dynamically varied so that data items that are frequently
updated are polled more often than data items that are rela-
tively static.
The home regions assign a TTR value to each of their
data items. This TTR value is varied dynamically by the
home region to reflect the update rate for the associated
data item. When the home region for data item d receives
an update for d, it locates the peer Pi which holds d. If
tupd intvl is the interval between successive updates of d,
and α (whose value is between 0 and 1 is a constant fac-
tor to weigh the relative importance of the recent and past
updates), then the TTR value is updated as:
T T R = α × T T R + (1 − α) × tupd intvl
This TTR value will be sent by the home region to other
peers on subsequent requests for data item d. When a peer
needs d and there is a local cached copy, it first determines
if the TTR of d has expired. The peer polls the home region
only if the TTR of d has expired. Thus we avoid too many
polls to the home regions while still ensuring good consis-
tency. The detailed algorithm of pull phase is given in Fig-
ure 3.
Notations:
d: requested data item;
Preq : peer requesting d;
Rh : home region for data item d;
Procedure Pull
Begin
When Preq requests data item d
if ( d is cached in Preq ) then
if ( TTR for d has expired) then
Obtain the home region Rh for d
Poll the home region Rh
Inquire for missed updates based on the last update times
Update the TTR and local cached copy
else
Use the cached copy of d to satisfy the request
End
Figure 3. Pull Phase
The main advantages of our Push with Adaptive Pull
mechanism are: 1) The push messages need to be sent only
to the home and replica regions, and not to the entire net-
work, thus saving bandwidth and power consumption; 2)
As the frequency of polling is determined by the update
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
rate for the data item, redundant polls to the home region
are avoided. This saves the bandwidth as well as reduces la-
tency.
5. Performance Analysis of PReCinCt
In this section we describe the energy model used in
our performance evaluation and determine the energy con-
sumption of the flooding scheme and proposed PReCinCt
scheme.
Energy as a distributed network resource has unique
properties that distinguish it from other resources [6]. En-
ergy is non-renewable: a mobile node has a finite, mono-
tonically decreasing energy store. Unlike bandwidth the
energy cost requires separate calculations with respect to
the sender, the intended receiver and other nodes which
overhear the message. This makes it necessary to distin-
guish between broadcast traffic which is processed by all
nearby nodes and point-to-point traffic which is processed
by the intended receiver(s) and discarded by all other nearby
nodes.
(2)
5.1. Energy Model
According to [6], the energy consumed by a mobile peer
for sending, receiving or discarding a message is given by
the following linear equation:
Cost = m × size + b (3)
where size is the message size, m denotes the incremental
energy cost associated with a message, and b is the energy
cost for the overhead of message. The parameters m and
b are different for sending and receiving. In a majority of
mobile devices like laptops, palmtops and cellular phones,
communication is one of the major sources of energy con-
sumption that reduces battery life. Therefore, we only con-
sider the energy cost for data communication in our fol-
lowing analysis. In particular, we analyze energy consump-
tion for broadcast traffic and point-to-point traffic and then
derive equations for energy consumption for the flooding
scheme and our proposed PReCinCt scheme.
5.1.1. Broadcast Traffic Before sending a broadcast
packet, the sender listens briefly to the channel. If the chan-
nel is clear, the message is sent and received by all nodes in
the wireless (radio) range. Otherwise, the sender must back
off and try later. The cost associated with sending a broad-
cast packet is given by:
Ebd sd = mbd sd × size + bbd sd (4)
The cost associated in receiving a broadcast packet is:
Ebd rv = mbd rv × size + bbd rv (5)
 When a packet is broadcast, all nodes within the transmis-
sion range of the sender will receive it. Let N be the num-
ber of nodes in the network, and A be the service area. The
number of nodes in a square area is given by the node den-
sity δ,
δ = N/A
Let r be the transmission range. The area within the trans-
mission range of the sender is given by π × r2. Thus,
the number of nodes (within the transmission range of the
sender) receiving the broadcast packet is,
ζ = δ × π × r2
5.2.2. Energy Cost of the PReCinCt Scheme In
PReCinCt, we assume there is no dynamic cache space
at each node, when a node requests for a data item, it
uses its hash function to get the home region for the re-
quested data item. The node then sends the request to that
(6) region via intermediate nodes. After reaching the home re-
gion, localized flooding is used to locate the node with that
data. The response is sent back to the requesting node via
one or more intermediate nodes. If I is the number of in-
termediate nodes between the requesting and responding
nodes and n is the average number of nodes in a re-
gion, the energy cost of each request in PReCinCt is given
(7) by,

Hence the total cost associated with a broadcast send and EP ReCinCt = I × (Ep2p sd + Ep2p rv )

receive is, + EF looding in Region

+ I × (Ep2p sd + Ep2p rv ) (12)
Etotal bd = Ebd sd + ζ × Ebd rv (8)
The term EF looding in Region is the cost of flooding in a re-
gion and depends on the number of nodes n in the region.
5.1.2. Point-to-Point Traffic In this case, the fixed
Thus,
costs include both the channel access cost and the cost
due to medium access control (MAC) layer negotia- EP ReCinCt = I × (Ep2p sd + Ep2p rv )
tions [6]. Therefore, the cost of a point-to-point sending is + n × Etotal bd
given by,, + I × (Ep2p sd + Ep2p rv ) (13)
Ep2p sd (size) = mp2p sd × size + bp2p sd (9)
6. Simulation Results
The cost of a point-to-point receive is, To measure the performance of our data retrieval and
caching schemes, we simulated the algorithms on a variety
Ep2p rv (size) = mp2p rv × size + bp2p rv (10) of static and mobile network topologies. We focus mainly
on the mobile simulation results in this paper, as it is more
5.2. Energy Consumption Analysis of the demanding and challenging in MP2P networks. GPSR [8] is
used as the wireless routing protocol. We have modified it to
PReCinCt and Flooding Schemes
route to regions instead of specific destinations by forward-
ing the packet towards the center of the region and using
In the light of the discussions above, we now derive
broadcast inside the region. The performance metrics mea-
expressions for energy consumption of the flooding and
sured were average latency and energy consumption. We
PReCInCt scheme.
compare our cache replacement algorithm GD-LD with the
well known GD-Size [2]. Likewise, we evaluate our cache
5.2.1. Energy Cost of the Flooding Scheme In the flood- consistency scheme namely Push with Adaptive Pull with
ing scheme as discussed in Section 1, when a node receives the Plain-Push and Pull-Every-time schemes.
a request for data, it broadcasts the request to all its neigh-
bors which in turn broadcast to each of their neighbors and
6.1. Simulation Environment
so on. The request will be processed by all the nodes in the
network. When the data is found, it is sent back to the re- We simulated our experimental setup in NS-2 [18]. The
questing node via a point to point link through one or more NS-2 simulation model simulates nodes moving in an unob-
intermediate nodes. If I is the number of intermediate nodes structed plane. Motion follows the random waypoint model
between the requesting and the responding node, and N is [1]. In our simulations the nodes are initially placed in a
the number of nodes in the network, the cost of the Flood- rectangular region of 1200m × 1200m divided into equal
ing scheme is given by, sized regions. We conducted simulations for a network of
up to 160 nodes with a nominal 250m transmission range
and a wireless bandwidth of 11Mbps. The time interval be-
EF looding = N ×Etotal bd +I ×(Ep2p sd +Ep2p rv ) (11) tween two consecutive requests and updates generated from

Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
 a peer follows a Poisson distribution with a mean of 30 sec- 0.65
GD−Size
onds. Each peer generates accesses to data items following 0.64
GD− LD

a Zipf distribution with a skewness parameter, Θ. Nodes

pause for a period of 5s between motion steps. We simu- 0.63

late maximum velocities of 2, 8, 12, 16 and 20m/s. The de-

Latency/Request(s)
0.62

fault number of regions is 9. 0.61

0.6

6.2. Simulation Experiments 0.59

0.58

The PReCinCt scheme is compared with the flooding

0.57
and the expanding ring search schemes for energy con- 0.5 1 1.5

Cache Size(% of Database Size)

2 2.5

sumption under varying node densities and moving speeds

in [11]. Three sets of experiments are reported in this pa- Figure 4. Variation of latency with cache size
per for evaluating the cooperative caching and consistency
mechanisms of PReCinCt. In the first set of experiments
our GD-LD cache replacement algorithm is compared with 0.5
GD−Size
GD−LD
the GD-Size algorithm. The performance metrics measured
0.45
were byte hit ratio and latency. The second set of experi-
ments compare different cache consistency algorithms and 0.4

show that our proposed Push with Adaptive Pull scheme

Byte Hit Ratio

performs better. Finally the third set of experiments vali- 0.35

dates the theoretical studies by comparing the theoretical

0.3

and simulation results.

0.25

6.2.1. Cache Replacement Experiments In these set of

experiments the developed cache replacement algorithm 0.2
0.5 1 1.5 2 2.5

Cache Size(% of Database Size)

GD-LD discussed in Section 2 is compared with GD-Size
[2]. The simulations were conducted on a topology with 80
nodes moving at a speed of 6m/s.
Figure 4 shows the variation of latency to fetch data
items with varying cache sizes. GD-LD by far outperforms
the GD-Size algorithm for all cache sizes. This is due to two
reasons: (1) The developed algorithm includes a new pa-
rameter - region distance while calculating the utility value
for an item. Region distance is the distance between the
regions of the requesting and responding peers. The algo-
rithm favors items that are further away as caching these
items would reduce latency and save bandwidth for subse-
quent requests for the same item, and (2) The GD-Size al-
gorithm, penalizes a large sized data item without consider-
ing its popularity or the cost of fetching it again for a sub-
sequent request.
Figure 5 shows the byte hit ratio of the two caching algo-
rithms. We observe GD-LD is able to achieve much higher
byte hit ratios as compared to those with GD-Size. This is
because GD-Size favors small data items independent of
their popularity, thus a large popular data item stands less
chance of being cached under GD-Size.
6.2.2. Data Consistency Experiments In this section we
compare our Push with Adaptive Pull cache consistency al-
gorithm with the Plain-Push and Pull-Every-time schemes.
The performance metrics measured are: 1) Control Message
Figure 5. Variation of byte hit ratio with cache
size
Overhead, which is measured as the total number of mes-
sages generated in the network to maintain data consistency
among the replicas; 2) False Hit ratio(FHR), measured as
the ratio of the number of stale hits to the number of hits
that are shown as valid; and 3) The average latency to re-
trieve a data item.
In our model, the time interval between two consecutive
requests, Trequest and the time between two consecutive up-
dates, Tupdate generated from a peer follows a Poisson dis-
tribution. We measure the effects of the time between suc-
cessive updates Tupdate on the different cache consistency
schemes. We fix Trequest to 30 seconds, and vary Tupdate so
that the ratio of Tupdate /Trequest varies from 1 to 5. A ra-
tio of 1 indicates the highest update rate.
Figure 6 shows the control message overhead of the
three consistency schemes. The message overhead of all
the schemes decreases with the update rate due to lesser
invalidation messages that are generated. The overhead of
Plain-Push is extremely high as the invalidation messages
are flooded to the entire network. In our Push with Adaptive
Pull scheme the invalidation messages are pushed only to
the corresponding home region thus incurring almost 89%

Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
 6 0.9
10
Pull−Every−time Pull−Every−time
Push−with−Adaptive−Pull Push−With−Adaptive−Pull
Plain−Push 0.85 Plain−Push

0.8

Control Message Overhead in log scale

5
10 0.75

Latency/Request(s)
0.7

0.65

4
10 0.6

0.55

0.5

3
10 0.45
1 1.5 2 2.5 3 3.5 4 4.5 5 1 1.5 2 2.5 3 3.5 4 4.5 5
Tupdate / Trequest
T /T
update request

Figure 6. Effect of update rate on the control Figure 8. Effect of update rate on the latency
message overhead per request

less message overhead than the Plain-Push scheme. The Figure 8 shows the average latency to retrieve a data item
Pull-Every-time scheme also incurs higher overhead than with varying update rates. It is observed that the Pull-Every-
that of our scheme, as in the Pull-Every-time scheme the time scheme has the highest average latency, as the peers are
peers are required to poll the home regions for every data required to poll the home regions for every request, thus in-
request to check for the validity of their cached item. In curring an extra round-trip delay in obtaining the requested
contrast, our scheme polls the home region only when the data.
TTR of that data item has expired, thus achieving 24%-57%
less overhead than Pull- Every-time. As the update rate in- 6.2.3. Validation of Theoretical Analysis In this section
creases our adaptive scheme tends to produce smaller TTR we validate the theoretical studies by comparing the results
values, resulting in slightly more polls. obtained from the theoretical analysis with those obtained
The FHR of the three schemes is shown in Figure 7. from the simulation experiments. The results are for a static
We observe that our scheme incurs the highest FHR, as the topology with an area of 600m × 600m.
peers poll the home regions for data only when the TTR
has expired, thus increasing the probability of a false hit.
300 200
However, we see that this ratio is very small, 0.01 even PReCinCt Theoretical PReCinCt Theoretical
250 PreCinCt Simulation PReCinCt Simulation
with highest update rate. ThePlain-Push scheme also in-
Energy/Request(mJ)

Energy/Request(mJ)
Flooding Theoretical 150
200 Flooding Simulation
curs some false hits, as it is possible that the invalidation
messages do not reach all the peers due to network conges- 150 100

tion, network partition or peer mobility. 100

50
50

0 0
0.01 20 40 60 80 0 5 10 15 20 25
Pull−Every−time No.of Nodes No. of Regions
Push−With−Adaptive−Pull
0.009 Plain−Push
(a) (b)

0.008
Figure 9. Comparison of theoretical and sim-
0.007
ulation results
0.006
False Hit Ratio

0.005

0.004

0.003 Figure 9(a) shows the theoretical and simulation results

0.002 for the energy consumed per request for the flooding and
0.001 PReCinCt schemes. We see that the simulation and theo-
0
1 1.5 2 2.5 3 3.5 4 4.5 5
retical results match when the node densities are small, but
Tupdate / Trequest
the difference between the results increases with larger node
Figure 7. Effect of update rate on the false hit densities. This difference is due to the well known ’edge-
ratio effects’ in flooding. When nodes at the edges of the topol-
ogy broadcast a message, the message is received by fewer
than expected number of nodes. However the edge effects

Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
 in our scheme are not as pronounced as those in the flood-
ing scheme.
Figure 9(b) shows the performance of our scheme with
varying number of regions. In this experiment the number
of nodes is set to 20. The number of regions is an impor-
tant design parameter that indicates the region size. Ide-
ally we would like to have regions with small size to reduce
the flooding overhead within a region and at the same time
to have a region size large enough to reduce the effects of
mobility. We observe that the scheme performs better and
consumes lesser energy with larger number of regions be-
cause the flooding takes place in smaller regions. Flooding
within smaller regions reduces the number of nodes pro-
cessing each query. Thus the energy consumed decreases
with increasing number of regions.
7. Conclusions and Future Work
This paper presents a cooperative caching and data con-
sistency scheme for PReCinCt for MP2P networks.
PReCinCt incorporates a cooperative caching scheme to
improve the communication performance of MP2P net-
works. This cooperative caching scheme enables peers
in a region to share their data, thus providing a uni-
fied view of the cache. This helps in alleviating the message
latency and limited accessibility problems in MP2P net-
works. The caching scheme includes a cache admission
control policy and a cache replacement algorithm. In or-
der to handle dynamic data, PreCinCt also includes an
effective cache consistency scheme called Push with Adap-
tive Pull that incurs less control message overheads as
compared to that of existing schemes. PReCInCt em-
ploys replication to achieve fault-tolerance. We also discuss
the management of regions in PReCInCt.
Future work includes an exhaustive analytical and ex-
perimental investigation on the impact of region size on our
scheme. A dynamic region management scheme need to be
investigated to make PReCinCt adaptive to real network en-
vironments, therefor optimizing its performance. In our fu-
ture work, various experiments need to be conducted to ver-
ify the robust performance of PReCinCt scheme under dif-
ferent mobility models and node disconnection rates.
References
[1] J. Broch, D. Maltz, D. Johnson, Y. Hu, and J. Jetcheva, A per-
formance comparison of multi-hop wireless ad hoc network
routing protocols In Proc. of the ACM/IEEE International
Conference on Mobile Computing and Networking (Mobi-
Com’98), August 1998.
[2] P. Cao, and S. Irani, Cost-aware WWW proxy caching al-
gorithms In Proceedings of USENIX Symposium on Internet
Technology and Systems, December, 1997.
Proceedings of the 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05)
1530-2075/05 $ 20.00 IEEE
[3] P. Cao and C. Liu, Maintaining Strong Cache consistency
in the World Wide Web, IEEE Transactions on Computers,
volume. 47, no. 4, pp. 445-457, Apr. 1998.
[4] A. Datta, M. Hauswirth, and K. Aberer, Updates in Highly
Unreliable, Replicated Peer-to-Peer Systems, In Proceed-
ings of ICDCS 2003, 23rd International Conference on Dis-
tributed Computing Systems, Providence, Rhode Island, May
2003, pp. 76C85.
[5] L. Fan, Summary Cache: A Scalable Wide-Area Web Cache
Sharing Protocol. In Proceedings of ACM SIGCOMM Conf.,
ACM Press, 1998
[6] L. M. Feeney, An energy consumption model for perfor-
mance analysis of routing protocols for mobile ad hoc net-
works. In ACM Journal on Mobile Networks and Applica-
tions 6, 2001.
[7] J. Gwertzman and M. Seltzer, World-Wide Web Cache Con-
sistency, In Proceedings of the 1996 USENIX Technical Con-
ference,January 1996.
[8] B. Karp, and H. T. Kung, GPSR: Greedy Perimeter Stateless
Routing for Wireless Networks In Proc. of the ACM/IEEE
International Conference on Mobile Computing and Net-
working (Mobicom 2000), August 2000.
[9] A. Kahol, S. Khurana, S.K.S. Gupta and P.K. Srimani, A
strategy to manage cache consistency in a distributed mo-
bile wireless environment, IEEE Transactions on Parallel
and Distributed System, 12(7), pp 686- 700, 2001.
[10] G. Kortuem, J. Schneider, D. Preuitt, T.G.C. Thompson, S.
Fickas, and Z. Segall, When peer-to-peer comes face-to-face:
collaborative peer-to-peer computing in mobile ad-hoc net-
works, In Proceedings of the First International Conference
on Peer-to-Peer Computing,August 2001.
[11] M.S. Joesph, M. Kumar, H. Shen and S.K. Das, Energy Ef-
ficient Data Retrieval and Caching in Mobile Peer-to-Peer
Networks , Workshop on Mobile Peer-to- Peer Computing,
Third International Conference on Pervasive Computig and
Communications, Kauai, Hawaii, March 8-12, 2005.
[12] Q. Lv, P. Cao, E. Cohen, K. Li, and S. Shenker, Search and
replication in unstructured peer-to-peer networks In Pro-
ceedings of the 16th annual ACM International Conference
on Supercomputing, 2002.
[13] P. Sarkar, and J. H. Hartman, Hint-based cooperative caching
ACM Transactions on Computer Systems, volume 18, num-
ber 4, pages 387–419, 2000.
[14] H. Shen, M. Kumar, S.K. Das, and Z. Wang, Energy-
Efficient Caching and Prefetching with Data Consistency in
Mobile Distributed Systems In Proc. of IEEE International
Parallel and Distributed Processing Symposium (IPDPS),
Santa Fe, NM, April 2004.
[15] I. Stoica, R. Morris, D. Karger, F. Kaashoek, and H. Balakr-
ishnan, Chord: A Scalable Peer-ot-Peer Lookup Services for
Internet Applications In Proc. of ACM SIGCOMM, 2001.
[16] K.L. Wu, P.S. Yu and M.S. Chen, Energy-efficient caching
for wireless mobile computing, In 20th International Con-
ference on Data Engineering, pp 336-345, 1996
[17] The Gnutella Homepage. http://gnutella.wego.com/.
[18] The VINT Project. The UCB/LBNL/VINT Network
Simulator-ns(version 2).http://mash.cs.berkeley.edu/ns.