Ad-hoc Storage Overlay System (ASOS): A Delay-Tolerant Approach in MANETs

Page created by Kathryn Robertson
 
CONTINUE READING
Ad-hoc Storage Overlay System (ASOS): A Delay-Tolerant Approach in MANETs
Ad-hoc Storage Overlay System (ASOS):
          A Delay-Tolerant Approach in MANETs
                   Guang Yang1 , Ling-Jyh Chen2 , Tony Sun1 , Biao Zhou1 and Mario Gerla1
                                    1
                                      Computer Science Department, UCLA
                                   {yangg, tonysun, zhb, gerla}@cs.ucla.edu
                        2
                          Institute of Information Science, Academia Sinica, Taiwan
                                           {cclljj}@iis.sinica.edu.tw

      Abstract— Mobile Ad-hoc NETworks (MANETs)                       quickly. On the other hand, MANETs often oper-
   are most useful in unprepared emergencies where crit-              ate in an adverse environment and are far less reli-
   ical applications must be launched quickly. However,               able than wired or infrastructure-based wireless net-
   they often operate in an adverse environment where                 works. Nodes in a MANET can crash, lose power,
   end-to-end connectivity is highly susceptible to disrup-
                                                                      be blocked, or move out of the communication range
   tion. Adjusting the motion of existing nodes or deploy-
   ing additional nodes can improve the connectivity un-              of its neighbors, etc. It is very difficult to guarantee
   der some circumstances, but for scenarios where con-               continuous end-to-end connectivity in MANETs.
   nectivity cannot be immediately improved, disruption                  Researchers have been looking at maintaining
   must be coped with properly. In this paper we propose              connectivity in MANETs for years. For exam-
   the Ad-hoc Storage Overlay System (ASOS). ASOS is                  ple, routing protocols employ redundancy and lo-
   a self-organized overlay of storage-abundant nodes to
                                                                      cal repair to minimize the chance of path break-
   jointly provide distributed and reliable storage to data
   flows under disruption. ASOS is a Delay-Tolerant Net-               age [11][17]. This alleviates the problem by find-
   working (DTN) approach that significantly improves                  ing an alternative path faster but only when such
   the applicability of MANETs in practice.                           a path exists. Others have looked at bridging dis-
     Index Terms—Mobile Ad-hoc Networks, Delay Tol-
                                                                      connected network partitions with additional nodes.
   erant Networking, Overlay, Storage                                 [14] studied how to reconnect a partitioned network
                                                                      by deploying as few additional nodes as possible.
                                                                      [23][27] proposed using additional mobile nodes to
                       I. I NTRODUCTION
                                                                      relay messages between network partitions. These
      Mobile Ad-hoc NETworks (MANETs) are a type                      methods, when applicable, can improve connectiv-
   of wireless networks that can be set up rapidly                    ity fast. However, there still exist situations where
   without pre-deployed infrastructure. They are ideal                additional nodes cannot be deployed as needed.
   where timely deployment of network infrastructure
                                                                         We believe disruption should be properly toler-
   is impractical, e.g. on a military battlefield or in dis-
                                                                      ated if it is unavoidable. Simple methods could
   aster recovery. Vehicular networks are also amongst
                                                                      save the data at the source node or some centralized
   the recent advances in MANETs that provide flexible
                                                                      server. They would obviously suffer various scala-
   data communications as alternative to infrastructure-
                                                                      bility and robustness issues. A more sophisticated
   based services. MANETs are most useful in un-
                                                                      scheme should involve replication and distributed
   prepared situations where tasks must be fulfilled
                                                                      storage for better reliability and performance. In our
      This material is based upon work supported in part by the Na-   proposed solution, named the Ad-hoc Storage Over-
   tional Science Foundation (NSF) under Grant No. 0520332, and       lay System (ASOS), storage is handled by a peer-to-
   in part by the Office of Naval Research (ONR) under Grant No.
   N00014-01-C-0016. Any opinions, findings, and conclusions or
                                                                      peer (P2P) overlay of memory-abundant nodes that
   recommendations expressed in this material are those of the au-    are part of the original MANET. These nodes, called
   thor(s) and do not necessarily reflect the views of NSF or ONR.     ASOS peers, jointly provide reliable data storage un-

1-4244-0507-6/06/$20.00 ©2006 IEEE                                296
der disruption. When an end-to-end flow is affected,        B. Design Principles
ASOS receives data from the source node, stores it            ASOS aims to extend the end-to-end transport
across the overlay, and delivers it to the original des-   model with the following design principles:
tination when connectivity is restored. ASOS is self-         P RINCIPLE 1: safe and robust storage. Data must
organized and complements the aforementioned ex-           be stored in a distributed manner with redundancy,
isting approaches. It is a Delay Tolerant Network-         and survive small-scaled failures.
ing (DTN) approach [4] specifically designed for               P RINCIPLE 2: immediate availability. ASOS
MANETs, and can be integrated as part of the DTN           must solely rely on the collaboration of existing
reference framework [4].                                   nodes and be available as soon as disruption occurs.
   The rest of the paper is organized as follows. Sec-        P RINCIPLE 3: efficient storage and easy delivery.
tion II illustrates application scenarios of ASOS and      ASOS storage should be managed efficiently and fa-
discusses its design principles. Section III proposes      cilitate future data delivery.
the ASOS design, focusing on overlay maintenance              P RINCIPLE 4: friendly interface. ASOS should
and service interfacing. Probabilistic data replication    provide a friendly interface of its storage utility both
is studied in Section IV under our mobility model.         source and destination nodes.
Section V presents simulation results to assess the
efficacy and performance of ASOS. Related work is                    III. T HE ASOS A RCHITECTURE
summarized in Section VI. Finally Section VII con-         A. Initialization and Maintenance of ASOS
cludes the paper.

    II. A PPLICATION S CENARIOS AND D ESIGN
                     P RINCIPLES
A. Application Scenarios
   ASOS is designed for Delay-Tolerant applications
running in a heterogenous MANET environment.
Most conventional applications such as file trans-
fer are considered as delay tolerant; some multime-
dia applications are also (partially) delay tolerant as
illustrated shortly. Heterogeneity means although
some nodes in the MANET may operate with limited
                                                           Fig. 1. ASOS architecture. When the destination is discon-
energy or storage, there exist powerful nodes with         nected from the source (1), data is submitted to ASOS for stor-
abundant resources, e.g. motor vehicles.                   age (1’). Stored data is delivered to the destination (2’) after it
   Assume a MANET is deployed for reconnais-               is reconnected (2).
sance in a large area. A number of mobile nodes,
humans and vehicles, are equipped with video cam-             1) Selecting ASOS peers: As illustrated in Fig-
eras. Distributed across the area, these nodes capture     ure 1, ASOS is a self-organized P2P overlay of ex-
useful data and send back to the control center. In the    isting nodes; every node in the MANET could be an
ideal case, all data is promptly received. In reality,     ASOS peer. Practically, it is desirable to designate
part of the data may not be delivered to the control       only a set of powerful nodes with abundant storage
center in time due to connectivity disruptions. The        as ASOS peers. We assume these nodes are pre-
control center can still reconstruct images from par-      configured. Other regular nodes must understand the
tially received data in real time at a reduced quality.    ASOS interface to access the storage utility.
ASOS will store the undeliverable data in a robust            2) Peer and file IDs: A number of P2P sys-
way, and deliver it to the control center later to re-     tems generate location-independent hash IDs for
construct the complete images. In this way ASOS            both peers and files; a file is stored at peers whose
does not improve the connectivity instantly, but pro-      IDs best match the file ID [21][24]. This method au-
vides a mechanism to save data that will eventually        tomatically spreads files uniformly across geograph-
contribute to the usefulness of the application.           ical locations, a desirable feature for large-scale file

                                                       297
sharing. This uniformity, however, is undesirable                  larger than the refresh interval of HELLO messages.
in ASOS, of which the primary goal is to deliver                   Expiration of a timer means that HELLO messages
a file1 to its original destination. We need an ID-                 have not been heard from the particular ASOS peer
independent algorithm to select storage locations for              recently, indicating the peer may have become un-
data, which will be studied in detail later in this pa-            reachable. The associated entry is then deleted from
per. For now, we assume every node has a unique                    the lookup table.
ID. Files can be uniquely identified by hashing from
source/destination IDs and other information.                      B. ASOS Interface
   3) Initialization and maintenance: ASOS is ini-                    1) Advertising of ASOS peers: Regular nodes in
tialized shortly after deployment. All ASOS peers                  a MANET must know the existence of nearby ASOS
form a multicast group, of which the address is                    peers. Due to the broadcast nature of wireless me-
known a priori by all peers. Discussion on alterna-                dia, the HELLO messages exchanged between ASOS
tive methods is beyond the scope of this paper. Each               peers could also be used for advertising. By enabling
ASOS peer multicasts periodic HELLO messages to                    the promiscuous mode, regular nodes could look
initialize and maintain the overlay. HELLO messages                at the HELLO messages to learn about the nearby
are sent to all ASOS peers, so that every peer can                 ASOS peers and stored files. Overhead is a seri-
hear from all other reachable peers and know which                 ous drawback of this scheme, in which HELLO mes-
files are stored in ASOS and where they are.                        sages from all ASOS peers must be sent to all reg-
                                                                   ular nodes. Aggregation on HELLO messages could
                                                                   alleviate the problem, but then would intervene with
                                                                   ASOS maintenance. We decide to decouple the task
                                                                   of advertising from maintenance, and introduce a
                                                                   new type of ADVERTISE messages.
                                                                      Each ASOS peer periodically aggregates its
                                                                   knowledge of reachable peers and stored files, and
                                                                   broadcasts an ADVERTISE message to all nodes.
                                                                   Since regular nodes do not participate in ASOS
Fig. 2. Format of a HELLO message.
                                                                   maintenance, fields such as remaining capacity and
                                                                   peer location are not included in ADVERTISE mes-
   Essential fields of a HELLO message is illustrated               sages. More importantly, ADVERTISE messages
in Figure 2. First, seqno is a peer-specific sequence               only contain file IDs rather than the detailed loca-
number incremented every time a new HELLO mes-                     tions of each file. This significantly reduces the size
sage is sent. A new message with a higher se-                      of the message. To suppress excessive broadcasting,
quence number supersedes previous messages from                    one node only forwards the first fresh ADVERTISE
the same peer. Remaining capacity and peer location                message it receives. This also guarantees that a reg-
are used for data management in ASOS. Stored files                  ular node receives ADVERTISE messages from its
are the meta data of the files, e.g. the file ID, start              closest ASOS peer.
offset, file size, etc. We will explain them shortly.                  2) Disruption detection and data submission:
   To keep track of peers and files, each ASOS peer                 We design ASOS as a backup scheme when end-to-
maintains a lookup table of reachable neighbors.                   end connectivity is disrupted. Disruption is usually
Entries in the table contain similar fields as in the               detectable at the routing layer. For example, popu-
HELLO messages. Each entry corresponds to one                      lar ad-hoc routing protocols [11][17] use RERR mes-
peer and is refreshed when a new HELLO message                     sages to report route errors. Receipt of an error mes-
from that peer is heard. Each entry is also associ-                sage is good indication of disruptions in connectiv-
ated with an expiration timer that is reset when the               ity. Note that many routing protocols perform local
entry is refreshed. The initial value of the timer is              repair before notifying the application. ASOS ac-
  1
  For simplicity, hereafter in this paper we assume that data in   commodates such efforts and will only intervene af-
ASOS is managed as files.                                           ter local repair fails.

                                                               298
files for it. The push mode is applicable when an
                                                                  original destination node is seen by an ASOS peer,
                                                                  e.g. appearing in the peer’s routing table.

                                                                  C. Data Management
                                                                     1) Probabilistic selection of storage locations:
                                                                  After the source node submits data to its ASOS
                                                                  agent, the agent becomes the first ASOS peer to hold
                                                                  a copy of the data. To increase storage reliability,
                                                                  data is also replicated to other ASOS peers. In-
                                                                  tuitively, it is desirable to store data near the des-
Fig. 3. Two methods to switch an end-to-end flow to the ASOS
mode: by the source node or an intermediate ASOS peer.            tination. We assume that nodes in the MANET
                                                                  know their (relative) locations. If not available, geo-
                                                                  information can be substituted by other metrics such
   Upon disruption detection, there are two meth-                 as the hop count.
ods to switch an end-to-end flow to the ASOS
mode. First, the source node can initiate the switch-
ing by contacting its ASOS agent (known from
ADVERTISE messages). Alternatively, an ASOS
peer on the path of the flow, once seeing an RERR
message, can initiate the switching immediately. See
Figure 3 for an illustration.
   We treat the entire data of an end-to-end flow as a
file. Due to disruptions, transmission of the file may
be divided into several chunks. Therefore, the start              Fig. 5. Probabilistic replication of data. Node B has a lower
offset in the file and the size of the chunk are critical          but non-zero probability of holding a copy. Nodes C and D
                                                                  have comparable probabilities. None of them deterministically
information in the local lookup table. See Figure 4
                                                                  hold a copy.
for an illustration.

                                                                      A naive way to replicate data is to greedily push
                                                                  the data towards its destination. This method has a
                                                                  drawback: replicas will be geographically close to
                                                                  each other and vulnerable to clustered failures. More
                                                                  sophisticated is the probabilistic replication across
                                                                  the overlay. The ASOS agent selects among its
                                                                  reachable peer neighbors K − 1 locations to repli-
                                                                  cate the data to, K being a configurable parame-
Fig. 4. Distributed storage of a file in ASOS. The first 200 data   ter. With the assumption that pairwise distances be-
units of File 1 are delivered to the destination. The next 150    tween nodes can be measured, storage locations are
units are stored at an earlier ASOS agent A. The current ASOS     selected based on the following guidelines (also see
agent is B with 50 units stored.
                                                                  Figure 5):
                                                                     i) A peer closer to the destination node should
   3) Data retrieval from ASOS: In terms of which                       have a higher probability to be selected,
party initializes the retrieval process, data can be ei-            ii) A peer further away from other ASOS peers that
ther pushed to the destination node by ASOS, or                         have been selected as storage locations should
pulled by the destination node itself. The pull model                   have a higher probability to be selected, and
is applicable when a destination node, by receiving                iii) A peer less heavily loaded should have a higher
ADVERTISE messages, learns that ASOS has stored                         probability to be selected.

                                                              299
The issue of probabilistic data replication will be
studied in further detail in the next section.
   2) Other data transfer between ASOS peers:
In addition to the probabilistic data replication, an
ASOS peer may also dynamically transfer stored
data to another peer. This can happen when a peer
is running short of power or storage space. A peer
may replicate data when it detects that a former peer
holding a replica has recently failed. A third situa-
tion is when the design parameter K, the number of
                                                         Fig. 6. The Virtual Track mobility model.
copies of the data, needs to be increased.
   3) Data deletion and replacement: ASOS sup-
ports both implicit and explicit data deletion and
                                                         B. ASOS Peer Deployment and Probabilistic Loca-
replacement. In the explicit scheme, the original
                                                         tion Selection under the VT Mobility Model
source or destination deletes data from ASOS by
messaging its ASOS agent, when the data is suc-              All three types of nodes, i.e. static, individual and
cessfully delivered or has lost its usefulness. The      grouped, can be ASOS peers. Initially, each group
source node ID, destination node ID, and file ID are      contains a certain number of ASOS peers. These
required to identify the file to be removed. The agent    in-group peers respond fast when storage is needed.
disseminates this message to all ASOS peers that         Due to splitting, a group may temporarily have zero
hold a copy of the data, which will then delete the      ASOS peers. In this case regular nodes turn to ASOS
data from their local storage. In the implicit scheme,   peers in nearby groups, or to static/individual peers
ASOS can accommodate storage scarcity with pri-          within reach.
oritized storage management, such as the Least Re-           Under the VT model, it is difficult to predict fu-
cently Used (LRU) and First-in-First-out (FIFO) al-      ture connectivity or the distance between two nodes:
gorithms. ASOS removes data deemed less impor-           the number of possible paths grows exponentially as
tant to make storage space for more valuable data.       nodes traverse across switch stations. Our approach
                                                         is to use the current positions and velocities of ASOS
    IV. P ROBABILISTIC DATA R EPLICATION                 peers to determine the replication locations. These
                                                         values are disseminated in the latest HELLO mes-
   Probabilistic data replication in ASOS can be for-
                                                         sages. We assume that the current position and ve-
mulated as an optimization problem. Due to space
                                                         locity of the destination node are also known. From
limit we do not discuss the generic problem in this
                                                         such information, the next switch station where an
paper; please refer to [25] for more details. We now
                                                         ASOS peer or the destination node will arrive can be
turn to a specific mobility model for a concrete loca-
                                                         determined. We use the distance between two switch
tion selection algorithm.
                                                         stations where two nodes will arrive, respectively, as
A. Virtual Track Mobility Model                          the distance between these two nodes.
                                                             The probabilistic location selection algorithm is
   The Virtual Track (VT) mobility model [28] tar-
                                                         shown in Table I. The weight wp meets all three
gets the MANET scenario where mobility of the
                                                         guidelines proposed in the previous section.
grouped nodes is constrained. It defines a set of
“switch stations” and “tracks”. Grouped nodes can
                                                                             V. E VALUATION
only move on the tracks. Nodes belonging to the
same group have the same group velocity; each node       A. Simulation Setup
then has its own internal velocity with respect to the      We have implemented ASOS as an application
group. Groups may split and/or merge at a switch         module in QualNet [18]. We select the UCLA cam-
station. Other than grouped nodes, there are also        pus map as the basis of our simulation scenario.
static and individual nodes in the VT model. Figure      Seven campus landmarks are identified as switch sta-
6 illustrates the concepts of the VT model.              tions of the VT mobility model. A total of 30 nodes

                                                     300
a: ASOS agent
  z: original destination node
  P: set of reachable peers from a
  S: selected peers as storage locations, S ⊆ P
  K: number of data copies (assumed as pre-configured)

  STp : next switch station where p will arrive
  dp,q : distance between STp and STq
  cp : remaining capacity of peer p
  random(r1 , r2 ): uniform random number generator on [r1 , r2 )

  begin
    S←Φ
    if (|P| < K)
       S←P
    else
       repeat
         for each p ∈ P do
           dmin ← min{dp,q |q ∈ S}
           wp ← (cp · dmin )/dp,z
           rp ← random(0, wp )
         enddo
         S ← S ∪ {p|p ∈ P, rp = max{rq |q ∈ P}}                       Fig. 7. Screen snapshot of the simulation topology in QualNet.
         P ← P \ {p|p ∈ P, rp = max{rq |q ∈ P}}
       until (|S| = K − 1)
    endif
  end                                                                 B. Delivery Ratio
                                                                         We first compare the instantaneous throughput in
                           TABLE I
                                                                      non-ASOS and ASOS scenarios. Hereafter in this
      A LGORITHM OF PROBABILISTIC SELECTION OF
                                                                      paper, the non-ASOS scenario means data is al-
     REPLICATION LOCATIONS AMONG             ASOS PEERS .
                                                                      ways delivered end-to-end, while the ASOS scenario
                                                                      means data is normally delivered end-to-end but will
                                                                      switch to ASOS when connectivity is disrupted.
                                                                         Figure 8 shows the instantaneous throughput mea-
are deployed in the 1600 m × 1600 m square area: 5                    sured every minute at the destination node. Ideally,
static, 5 individual and 20 grouped. Mobility of indi-                this throughput should be 67 Kbps, the aggregate
vidual nodes follows the Random Way-Point (RWP)                       sending rate from all sources, for the first 10 minutes.
model; grouped nodes are divided into 4 groups of                     Due to disruption however, throughput in the non-
5 nodes each. The screen snapshot from QualNet is                     ASOS scenario is consistently below the ideal value.
shown in Figure 7. Nodes 1 to 5 are static, 6 to 10                   ASOS does not instantly improve connectivity, but
are individual, 11 to 30 are grouped.                                 temporarily stores data and resumes delivery when
   Node 30 is the destination where all data traf-                    connectivity is restored. Clearly reflected in the fig-
fic is directed. Five flows are set up from differ-                     ure, at the 8th minute the instantaneous throughput is
ent source nodes, including one static, one individ-                  well above 67 Kbps. This includes both fresh data
ual and three grouped nodes, none of which be-                        produced during the minute, and previously stored
long to the same group with node 30. Each source                      data. After the 10th minute when all flows have
generates a periodic constant-bit-rate (CBR) flow, at                  stopped, some data can still be delivered to the des-
the rate of 80 Kbps, for 10 seconds every minute.                     tination, e.g. at the 11th and 17th minute. Figure
Communication between nodes is IEEE 802.11b at                        9 shows the cumulative amount of data delivered to
2 M bps, with an effective transmission range of ap-                  the destination. At the end of the simulation, ASOS
proximately 280 m. Each simulation runs 20 min-                       is able to deliver about twice as much data as deliv-
utes. Data traffic stops at the 10th minute; ASOS, if                  ered in the non-ASOS scenario. We have repeated
enabled, has an additional 10 minutes to exploit node                 the simulation with different random seeds, seeing
mobility and deliver data to the destination. Simula-                 the similar trend.
tion parameters are summarized in Table II.                              Figure 10 shows how individual flows benefit

                                                                    301
topology                                                      1600m × 1600m                                              ASOS
                                                                                                                    5   Non ASOS
   simulation time                                               20 min
                                                                                                                             Ideal
                                                                 30

                                                                                          Delivered Data (MBytes)
   total number of nodes
   number of static nodes                                        5                                                  4
   number of individual nodes                                    5
   number of groups                                              4                                                  3
   number of nodes in each group                                 5
   number of static nodes as ASOS peers                          2                                                  2
   number of ind. nodes as ASOS peers                            2
   number of ASOS peers in each group                            2
                                                                                                                    1
   (excluding the group with node 30)
   number of replicas                                            3
                                                                                                                    0
                                                                                                                         2    4      6   8      10 12     14   16   18   20
   interval between HELLO messages                               10 sec
   interval between ADVERTISE messages                           10 sec                                                                      Time (min)
   expiration time of entries in lookup tables                   30 sec
                                                                                      Fig. 9. Cumulative amount of data delivered to the destination
   RWP model min speed                                           0 m/sec              as time proceeds.
   RWP model max speed                                           10 m/sec

   VT model group min speed                                      0 m/sec
   VT model group max speed                                      10 m/sec             proved the least. This is because static nodes lack
   VT model internal min speed                                   0 m/sec              the ability to move and actively meet other nodes.
   VT model internal max speed                                   1 m/sec
   VT model max track length                                     750 m                Individual nodes can move, but their territory can be
   VT model group split prob.                                    0.25                 much larger than the grouped nodes. Grouped nodes
   nominal 802.11b data rate                                     2 M bps              only move on the tracks; this effectively reduces
   approx. transmission range                                    280 m                their territory and increases the chance of meeting
   number of flows                                                5
   average per-flow data rate                                     13.33 Kbps
                                                                                      each other. Therefore, we have observed the best de-
   start time of flows                                            0th min              livery ratio improvement on grouped nodes.
   stop time of flows                                             10th min

                                              TABLE II                                C. Messaging Overhead
  S UMMARY OF SIMULATION PARAMETERS IN Q UAL N ET.
                                                                                         Both HELLO and ADVERTISE messages in
                                                                                      ASOS incur overhead, with HELLO messages being
                                                                                      the heavier source as they are generally larger. In
                           120                                                        Figure 11 we plot the cumulative distribution func-
                                                                ASOS
                                                            Non ASOS                  tion (CDF) of the size of HELLO messages. Over
                           100                                   Ideal
                                                                                      70% of the HELLO messages are less than 200 bytes
   Throughput (Kbps/sec)

                           80                                                         long, with a maximum of 460 bytes only. In our sim-
                                                                                      ulation, the control traffic is merely 1.6 Kbps, neg-
                           60
                                                                                      ligible compared to the data traffic.
                           40

                           20                                                         D. Impact of ASOS Parameters
                                                                                         Two key parameters in ASOS are the number of
                            0
                                 2   4    6    8     10   12    14   16   18   20     ASOS peers deployed in the MANET, and the num-
                                                   Time (min)
                                                                                      ber of data copies, i.e. K. So far they have been
Fig. 8.                     Instantaneous throughput measured at the destination      fixed at 10 and 3, respectively. We now vary these
node.                                                                                 numbers and study the impact on ASOS.
                                                                                         First we vary the number of ASOS peers from 5
                                                                                      to 20; the results are shown in Figure 12. At the be-
from ASOS: one from a static node (left), one from                                    ginning, the delivery ratio increases with the number
an individual node (center), and one from a grouped                                   of ASOS peers. This is obvious since more ASOS
node (right). The delivery ratio for the grouped node                                 peers provide better availability. However, the in-
has improved the most, while the static node has im-                                  creasing delivery ratio quickly reaches the peak and

                                                                                    302
1.2                                                                    1.2                                                                            1.2
                                                                 ASOS                                                                ASOS                                                                           ASOS
                                                             Non ASOS                                                            Non ASOS                                                                       Non ASOS
                                                        1         Ideal                                                     1         Ideal                                                                1         Ideal
                        Delivered Data (MBytes)

                                                                                                 Delivered Data (MBytes)

                                                                                                                                                                                Delivered Data (MBytes)
                                                    0.8                                                                    0.8                                                                            0.8

                                                    0.6                                                                    0.6                                                                            0.6

                                                    0.4                                                                    0.4                                                                            0.4

                                                    0.2                                                                    0.2                                                                            0.2

                                                        0                                                                   0                                                                              0
                                                             2    4   6   8 10 12 14 16 18 20                                    2    4    6   8 10 12 14 16 18 20                                              2   4   6    8 10 12 14 16 18 20
                                                                          Time (min)                                                           Time (min)                                                                    Time (min)

Fig. 10. Delivery ratios of data generated in each minute. Three flows are shown, one from a static node (left), one from an
individual node (center), one from a grouped node (right).

                                                   1                                                                                                We then fix the number of ASOS peers at 10 and
                                                            CDF
                                                                                                                                                 vary the number of copies from 1 to 5, results shown
                                                  0.8                                                                                            in Figure 13. At the beginning, the delivery ratio
                                                                                                                                                 grows with the number of copies, but quickly hits the
                                                  0.6
                                                                                                                                                 plateau; further increasing the number has little im-
   CDF

                                                  0.4
                                                                                                                                                 pact on the delivery ratio. This is different from what
                                                                                                                                                 we have observed in Figure 12. The reason is that
                                                  0.2                                                                                            increasing the number of ASOS peers leads to more
                                                                                                                                                 HELLO and ADVERTISE messages, while increas-
                                                   0                                                                                             ing the number of data copies leads to larger HELLO
                                                        0             100       200     300      400                                 500
                                                                      Size of HELLO messages (Bytes)                                             and ADVERTISE messages. Since these messages
                                                                                                                                                 are relatively small in our scenario, increase in the
Fig. 11. Cumulative distribution function (CDF) of the size of
HELLO messages.                                                                                                                                  number of messages has larger impact on the traffic
                                                                                                                                                 load than increase in the message size.

                                                                                                                                                                          100
then starts to decline as more ASOS peers are added.                                                                                                                                                                         Delivery Ratio

This is mainly due to the fact that excessive ASOS                                                                                                                        80
peers incur significantly more messaging overhead,
                                                                                                                                                     Delivery Ratio (%)

which negates the marginal gain brought by these                                                                                                                          60
additional ASOS peers.
                                                                                                                                                                          40

                                            100                                                                                                                           20
                                                                                      Delivery Ratio

                                                  80                                                                                                                        0
   Delivery Ratio (%)

                                                                                                                                                                                1                               2      3       4              5
                                                  60                                                                                                                                                            Number of Copies

                                                                                                                                                 Fig. 13. Delivery ratio vs. number of replicated data copies.
                                                  40
                                                                                                                                                 Deliver ratios are the overall values of all flows directed to the
                                                                                                                                                 destination.
                                                  20

                                                   0                                                                                               Figures 12 and 13 indicate that choosing the ap-
                                                                      5       10       15                        20
                                                                          Number of ASOS Peers
                                                                                                                                                 propriate number of ASOS peers and number of data
                                                                                                                                                 copies has big impact on the ASOS performance. It
Fig. 12. Delivery ratio vs. number of ASOS peers. Deliver ra-                                                                                    depends on a number of factors such as the topology,
tios are the overall values of all flows directed to the destination.
                                                                                                                                                 mobility and traffic patterns. We do not intend to ex-

                                                                                                                                               303
plore an optimal solution in this paper and leave it       and ad-hoc routing. Mobile Information Retrieval
for future study.                                          (IR) is studied in [7]. It splits, indexes and replicates
                                                           a given database probabilistically across all mobile
               VI. R ELATED W ORK                          nodes, and a node only contacts its 1-hop neigh-
   The DTN Research Group (DTNRG) [4] is ded-              bors to answer a query. PAN [15] is another prob-
icated on providing inter-operable communications          abilistic storage system in MANETs. In PAN, small
in challenged networks [5][9][10], which has en-           data objects are replicated to multiple servers and
lightened us on developing ASOS to cope with the           can be dynamically updated. To control overhead,
frequent connectivity disruptions in MANET scenar-         PAN “lazily” updates data such that inconsistency
ios. Among other DTN efforts for MANETs, [8]               may temporarily exist. It is difficult to apply Mo-
proposes using controlled flooding in sparse mobile         bile IR and PAN directly to ASOS for two reasons.
networks for packet delivery when end-to-end paths         First, the database in mobile IR is static, while in
do not exist. [13] studies several different oppor-        ASOS data is produced in real time. Second, both
tunistic forwarding strategies using mobile routers        mobile IR and PAN replicate data to improve chance
(similar to the mobile ferries [27]) in vehicular ad-      of access from all nodes; this is unnecessary and un-
hoc networks. Adaptive routing for intermittently          desirable in ASOS.
connected MANETs is investigated in [16]. Our                 Cooperative caching has also been studied for
work differs from these efforts and can be integrated      MANETs. [12] proposes adding an application man-
as an added feature in MANETs where the above              ager component between the network layer and ap-
schemes are already implemented.                           plications. The application manager will switch to
   P2P overlay storage is widely studied on the In-        a better data source if the QoS provided by the cur-
ternet. Cooperative File System (CFS) has been pro-        rent one is degrading. This requires that at least one
posed in [3]; PAST [20] is another P2P storage util-       data source is available at any time, which cannot
ity. These systems are good paradigms on P2P stor-         be guaranteed in our target scenarios. [22] and [26]
age; however, our target problem is different. CFS         have investigated mechanisms of caching popular
and PAST aim to provide fast file access to all po-         data among ad-hoc nodes, such that when the origi-
tential users on the Internet. Files are likely to be      nal source is not available, data can be obtained from
replicated geographically to many distant locations,       nearby nodes. These studies have revealed valuable
a desirable property for robust file sharing. On the        insight into distributed data storage in MANETs;
contrary, ASOS aims to extend the conventional end-        however, they do not target the problem of extend-
to-end delivery model where stored data is to be de-       ing end-to-end flows to survive disruptions.
livered to the intended original destination only.
   In sensor networks, Data Centric Storage (DCS)
                                                                            VII. C ONCLUSION
[19] is proposed to save all data of the same type
at a designated storage server, such that data queries        In this paper we have proposed the Ad-hoc Stor-
are directed only to the server. The basic DCS is          age Overlay System (ASOS) that extends end-to-end
extended with local refreshes and structured repli-        data transport in MANETs when connectivity is dis-
cation for better data robustness and system scala-        rupted. ASOS is a DTN approach to tolerate in-
bility. Resilient DCS (R-DCS) [6] further improves         evitable disruptions in MANETs, by storing undeliv-
the robustness and scalability by replicating data at      erable data reliably in an overlay of storage-abundant
strategic locations. DCS/R-DCS share many com-             nodes and delivering it later when connectivity is re-
monalities with the idea of ASOS; the main differ-         stored. We have implemented the ASOS in QualNet,
ences are: 1) DCS/R-DCS use geographic informa-            and shown that it can significantly increase the over-
tion mostly for routing while ASOS uses it for loca-       all data delivery ratio in disrupted MANET scenar-
tion selection, and 2) data replication is deterministic   ios.
in DCS/R-DCS but probabilistic in ASOS.                       In the future, we plan to investigate effective inte-
   In the context of MANETs, [2] has proposed to           gration of ASOS with the DTN reference implemen-
provide better interaction between P2P addressing          tation [4]. We also plan to address a few open issues.

                                                       304
First, we have only considered pre-configured ASOS                [12] W. Lau, M. Kumar and S. Venkatesh, “A Cooperative
peers in this paper. Alternatively ASOS peers can be                 Cache Architecture in Support of Caching Multimedia Ob-
                                                                     jects in MANETs”, ACM WoWMoM’02, Atlanta, GA, Sept
dynamically elected, e.g. when no ASOS peers are
                                                                     2002.
within the reach of regular nodes. Second, in order              [13] J. LeBrun, C. N. Chuah and D. Ghosal. “Knowledge Based
to better support multicast applications, the current                Opportunistic Forwarding in Vehicular Wireless Ad Hoc
data management and interface in ASOS need to be                     Networks”, IEEE VTC Spring 2005, Stockholm, Sweden,
                                                                     May 2005.
revised. Erasure codes can also be incorporated into
                                                                 [14] N. Li, and J. C. Hou, “Improving Connectivity of Wire-
ASOS to improve data reliability. Also, data encryp-                 less Ad-Hoc Networks”, UIUC DCS Technical Report
tion, user authentication and intrusion detection are                UIUCDCS-R-2004-2485, Oct 2004.
needed when ASOS operates in a hostile environ-                  [15] J. Luo, J-P Hubaux and P. T. Eugster, “PAN: Providing
ment against malicious attackers. Finally there exist                Reliable Storage in Mobile Ad Hoc Networks with Proba-
                                                                     bilistic Quorum Systems”, ACM MobiHoc 2003, Annapo-
“soft disruptions” where end-to-end connectivity is                  lis, MD, Jun 2003.
not totally lost but can be maintained at a reduced              [16] M. Musolesi, S. Hailes and C. Mascolo, “Adaptive Rout-
effective capacity. We are interested in upgrading                   ing for Intermittently Connected Mobile Ad Hoc Net-
ASOS for such scenarios.                                             works”, IEEE WoWMoM 2005, Taormina, Italy, Jun 2005.
                                                                 [17] C. E. Perkins, E. M. Belding-Royer and I. Chakeres,
                                                                     “Ad Hoc On Demand Distance Vector (AODV) Rout-
                                                                     ing”, IETF Internet draft, draft-perkins-manet-aodvbis-
                      R EFERENCES                                    00.txt, Oct 2003.
                                                                 [18] QualNet, http://www.qualnet.com/.
[1] S. Chessa and P. Maestrini, “Dependable and Secure Data      [19] S. Ratnasamy, D. Estrin, R. Govindan, B. Karp,
    Storage and Retrieval in Mobile, Wireless Networks”,             S. Shenker, L. Yin and F. Yu, “Data-Centric Stor-
    DSN’03, San Francisco, CA, Jun 2003.                             age in Sensornets”, http://lecs.cs.ucla.edu/
[2] M. Conti, E. Gregori and G. Turi, “Towards Scalable P2P          Publications/papers/dht.pdf.
    Computing for Mobile Ad Hoc Networks”, IEEE PerCom           [20] A. Rowstron and P. Druschel, “Storage management and
    Workshops 2004, Orlando, FL, Mar 2004.                           caching in PAST, a large-scale, persistent peer-to-peer stor-
[3] F. Dabek, M. F. Kaashoek, D. Karger, R. Morris and I.            age utility”, ACM SOSP 2001, Banff, Canada, Oct 2001.
    Stoica, “Wide-area cooperative storage with CFS”, ACM        [21] A. Rowstron and P. Druschel, “Pastry: Scalable, decentral-
    SOSP 2001, Banff, Canada, Oct 2001.                              ized object location and routing for large-scale peer-to-peer
[4] Delay Tolerant Networking Research Group, http://                systems”, IFIP/ACM Middleware 2001, Heidelberg, Ger-
    www.dtnrg.org.                                                   many, Nov 2001.
[5] K. Fall, “A Delay-Tolerant Network Architecture for Chal-    [22] F. Sailhan and V. Issarny, “Cooperative Caching in Ad
    lenged Internets”, ACM SIGCOMM 2003, Karlsruhe, Ger-             Hoc Networks”, The 4th International Conference on Mo-
    many, Aug 2003.                                                  bile Data Management (MDM), Jan 2003.
[6] A. Ghose, J. Grossklags and J. Chuang, “Resilient data-      [23] R. Shah, S. Roy, S. Jain and W. Brunette, “Data MULEs:
    centric storage in wireless ad-hoc sensor networks”, The         Modeling a Three-tier Architecture for Sparse Sensor Net-
    4th International Conference on Mobile Data Management           works”, Intel Research Technical Report, IRS-TR-03-001,
    (MDM), Jan 2003.                                                 Jan 2003.
[7] K. M. Hanna, B. N. Levine and R. Manmatha, “Mobile Dis-      [24] I. Stoica, R. Morris, D. Karger, M. F. Kaashoek and H.
    tributed Information Retrieval For Highly-Partitioned Net-       Balakrishnan, “Chord: A Scalable Peer-to-peer Lookup
    works”, IEEE ICNP 2003, Atlanta, GA, Nov 2003.                   Service for Internet Applications”, ACM SIGCOMM 2001,
[8] K. Harras, K. Almeroth and E. Belding-Royer, “Delay              San Diego, CA, Aug 2001.
    Tolerant Mobile Networks (DTMNs): Controlled Flood-          [25] G. Yang, L-J. Chen, T. Sun, B. Zhou and M. Gerla, “Ad-
    ing Schemes in Sparse Mobile Networks”, IFIP Netwoking           hoc Storage Overlay System (ASOS): A Delay-Tolerant
    2005, Waterloo, Canada, May 2005.                                Approach in MANETs”, UCLA CS Technical Report
[9] S. Jain, K. Fall and R. Patra, “Routing in a Delay Toler-        050030, Jul 2005.
    ant Networking”, ACM SIGCOMM 2004, Portland, OR,             [26] L. Yin and G. Cao, “Supporting Cooperative Caching
    Aug/Sept 2004.                                                   in Ad Hoc Networks”, IEEE Infocom 2004, Hong Kong,
[10] S. Jain, M. Demmer, R. Patra and K. Fall “Using Redun-          China, Mar 2004.
    dancy to Cope with Failures in a Delay Tolerant Network”,    [27] W. Zhao, M. Ammar and E. Zegura, “A Message Ferry-
    to appear in ACM SIGCOMM 2005, Philadelphia, PA, Aug             ing Approach for Data Delivery in Sparse Mobile Ad Hoc
    2005.                                                            Networks”, ACM MobiHoc 2004, Roippongi, Japan, May
[11] D. B. Johnson, D. A. Maltz and Y.-C. Hu, “The Dy-               2004.
    namic Source Routing Protocol for Mobile Ad Hoc Net-         [28] B. Zhou, K. Xu and M. Gerla, “Group and Swarm Mobil-
    works (DSR)”, IETF Internet draft, draft-ietf-manet-dsr-         ity Models For Ad Hoc Network Scenarios Using Virtual
    09.txt, Apr 2003.                                                Tracks”, MILCOM 2004, Monterey, CA, Oct/Nov 2004.

                                                             305
You can also read