IMPROVING THE LOCATION OF MOBILE AGENTS ON
SELF-AWARE NETWORKS
Ricardo Lent
Intelligent Systems and Networks, Imperial College London, U.K.
Keywords:
Mobile Agent Systems, Cognitive Packet Networks, Mesh Networking.
Abstract:
Business applications are evolving on the Internet from centralised to distributed architectures demanding
higher quality group communication facilities. Such communications are often time sensitive and prone to
disruptions or unacceptable latency because of the high variability of Internet traffic and capacity across the
network. Self-aware networks offer mesh networking facilities with adaptive routing able to cope with changes
in network conditions, such as variations in traffic load and link or node failure. Further quality improvement
can be achieved by relaying critical application tasks to software agents able to migrate to alternate hosts and
take advantage of new location facilities for their communications. The paper examines the concurrent use of
self-aware networking and software agent mobility to offer improved communication facilities to time-critical
group communications. A host selection algorithm for agent migration is proposed to find suitable locations
for agents and evaluated in a simulation study.
1 INTRODUCTION
A mobile agent is a type of software agent that can
autonomously migrate from one computer to another
to continue execution on the second host. Migration
involves moving the mobile agent’s code and data to-
tally or partially, and in some cases also the execution
state from one machine to another. Migration may of-
fer interesting advantages to critical business applica-
tions that would benefit from changing their execution
environment in response to a computing or commu-
nication degradation (Lange and Oshima, 1999). It
would be of particular use for time-constrained appli-
cations. Because of the great variability of resources
in the Internet, network performance in general can
vary greatly from one network area to another and
also over time. Significant network variability can in-
troduce an undesirable uncertainty on the quality of
the communications that may affect critical applica-
tions. Applications supported by mobile agents need
not to be bounded by a fixed execution environment
as they are able to relocate themselves and use a more
advantageous location offering better communication
quality.
For example, on a distributed auction system, auc-
tioneers need to notify bidders after they receive a
new offer to allow them react and send additional
bids. Un-even latency and packet loss in the commu-
nications between seller and bidders may cause un-
fair advantage to certain participants. Another case is
exemplified by the use of a central host to ensure to-
tal ordering in a chat system, whose location would
determine the performance of the application. Both
cases are good examples of critical real-time group
communications, that could benefit greatly from the
use of moving agents.
There are several comprehensive systems that sup-
port agent mobility that have been made available
both from industry and academia. From the point of
view of the form of mobility, those systems can be
classified into one of two categories: weak mobility
or strong mobility. The former form supports data
and code mobility whereas the latter supports in ad-
dition a transfer of the control state of the process.
The Java virtual machine natively supports weak mo-
bility so many mobile agent systems are Java-based.
Java offers portability with the use of bytecodes ex-
ecuted with just-in-time (JIT) compilation and host
security that is enforced by virtual machines. Aglets
(Tai and Kosaka, 1999), which have been developed
by IBM, Voyager (http://www.recursionsw.com/, ),
Concordia (Wong et al., 1997) and Jumping Beans
(http://www.JumpingBeans.com, ) are examples of
Java-based systems offering weak mobility. Tacoma
(Sudmann, 1996) and Concordia are systems offering
weak mobility. Tacoma offers support to multiple lan-
guages including Java. Some of the systems that offer
strong mobility are D’Agents (Gray, 1995; Gray et al.,
239
Lent R. (2008).
IMPROVING THE LOCATION OF MOBILE AGENTS ON SELF-AWARE NETWORKS.
In Proceedings of the Tenth International Conference on Enterprise Information Systems - SAIC, pages 239-242
DOI: 10.5220/0001700802390242
Copyright
c
 SciTePress

2002) and Ara (Peine, 2002) both of which supporting
multiple language support.
Self-aware networks are networks that exhibit a
number of autonomic properties, which are useful
to create effective mesh networks. These networks
can dynamically reconfigure paths, both to improve
communications quality and to avoid network ele-
ment failures. Cognitive packet networks (CPN) is a
self-aware network that utilizes reinforcement learn-
ing to dynamically adjust paths towards a user de-
fined QoS goal. The paper’s focus is close to the
work of (Ranganathan et al., 1999), who investigated
network-aware mobile programs, which are adaptable
to network variations. Ranganathan et al. proposed a
distributed latency monitor to probe the network and
verify its status. Because of its structure, this work
is applicable mainly to long-lasting communications
and coarse-grained network variations. Other migra-
tion strategies were evaluate by (Gonzalez-Valenzuela
and Vuong, 2002).
This paper tackles the problem of dynamically se-
lecting migration hosts for mobile agents for time-
critical group communications, which range from
short-lived to long-lived communications. The case
of short-lived communications is particularly chal-
lenging because network conditions tend to change
all the time, so that long-running network monitors or
static migration selection may not be effective. The
paper explores the use of a self-aware network as the
underlying communications infrastructure for agents,
which is able to provide critical information to guide
their migration. In particular, the paper uses CPN,
which when coupled with mobile agents can together
offer two levels of adaptation to applications. First,
the CPN provides self-adaptable routing for improved
quality of service (QoS). Second, the QoS of commu-
nications can potentially improve more by selecting
an appropriate node from which to host the commu-
nications.
2 SELECTION OF MIGRATION
HOST
Given a network represented by the graph G = (V, E),
where an agent located at node s has established
a time-critical communication with N other nodes:
t
1
, t
2
, . . . , t
N
and assuming that any node is able to host
mobile agents, the problem is to identify the most
suitable host v
∗
∈ V for a particular agent that mini-
mizes the cost of its group communication with nodes
t
1
, t
2
, . . . , t
N
. The solution would require knowledge of
the underlying network infrastructure along with an
observation of the network conditions.
The use of CPN or comparable self-aware net-
work as the supporting communications infrastructure
would allowboth route adaptation and to expose some
of the information that agents would need to make a
proper migration decision. In CPN, paths are con-
structed dynamically and updated over time. CPN
nodes can temporarily store a number of last discov-
ered paths (i.e. store the set Π
q
s,t
= {P
q
s,t
|P
q
s,t
is a path
from s to t built with routing goal q}). A path P
q
s,t
defines a sequence of nodes that dumb packets visit:
P
q
s,t
= v
1
, v
2
, . . . , v
k
, where k = |P
q
s,t
| is the length of the
path in number of hops and v
k
= t.
Assuming additive cost metrics, the cost of path
q(P
q
s,t
) is:
q(P
q
s,t
) =
k
∑
i=1
q(v
i
, v
i+1
)
where q : (v
i
, v
i+1
) → ℜ, so that q(v
i
, v
i+1
) is the
observed cost on edge (v
i
, v
i+1
). The observed cost of
the path is continuosly acquired by CPN packets.
Let Π
s
be the set of all paths known by node s:
Π
s
= {Π
q
s,t
|Π
q
s,t
is a set of paths from s to t built with
routing goal q} and G
′
= (V
′
, E
′
) be the inferred graph
from Π
s
by making V
′
= {v|v is a vertex of a path in
Π
s
} and E
′
= {e|e are two successive vertices of a
path in Π
s
}.
Likewise, let Q
s
be the function Q : (v
i
, v
j
) → ℜ
such that (v
i
, v
j
) ∈ E
′
and Q(v
i
, v
j
) = sel{q(v
i
, v
j
) :
(v
i
, v
j
) is an edge of a path in P
s
}, where sel is a func-
tion selecting the desired cost to work with, for exam-
ple: minimum, maximum, most recent, etc.
Group Distance. Φ : {u, s, t
1
, . . . , t
N
} → ℜ, such
that
Φ(u, s, t
1
, . . . , t
N
) =
v
u
u
t
c
0
|P
u,s
|
2
+
N
∑
i−1
c
i
|P
u,t
i
|
2
for a choice of constants c
i
, which weight the im-
portance of the path cost between the originating node
(s) and group members (t
i
). We will refer to s and
t
1
, . . . , t
k
as the reference nodes. For c
i
= 1, Φ repre-
sents the Euclidean distance from the selected migra-
tion host u to the reference nodes.
The problem is to find a node u in graph G
′
(and
ultimately in G) such that:
φ(u, s, t
1
, . . . , t
k
) = Φ(u, s, t
1
, . . . , t
N
) : u ∈ G
′
}
The value of Φ can be calculated from the com-
putation of the shortest path cost from each reference
ICEIS 2008 - International Conference on Enterprise Information Systems
240

node to allV
′
nodes. The computation has a complex-
ity of O(NE
′
+ NV
′
logE
′
) when done with the Di-
jkstra’s algorithm implemented with a priority queue
(Fibonacci heap).
Expected Trend. It is reasonable to expect that for
a large migration rate τ
−1
m
with respect to the group
communication lifetime and a supporting self-aware
network with reasonable active flows, within a finite
time:
φ(u, s, t
1
, . . . , t
k
) → φ(v
∗
, s, t
1
, . . . , t
k
)
that is, the agent will tend to move to the host in
G offering the minimum distance. The case is studied
in simulation in the next section.
3 EVALUATION
A simulation study was conducted to measure the po-
tential benefits of introducing a selection of the migra-
tion host for agents exploiting the information gath-
ered by a CPN. The simulation used INES (Lent,
2008).
The network topology in the study was a squared
grid of 16 nodes per side (256 nodes in total) and full-
duplex links with exponentially distributed propaga-
tion delays and transmission rates (with parameters
1 ms and 512 Kbps respectively). Each network in-
terface had a buffer capacity of 100 packets. In the
simulations, CPN was enabled to use external routing
information (i.e. from a non-dynamic routing algo-
rithm) to speed up the initial path setup.
The simulations used a 0.2 smart packet ratio with
a 0.01 probability of random selection. The maximum
number of hops was set to 64 and the routing goal of
smart packets was low latency. A full explanation of
these parameters can be found in the literature (Ge-
lenbe et al., 2002; Gelenbe et al., 2001).
The group communication application consisted
of a mobile agent sending traffic (10,000 packets of
256 bytes at 1 Kbps CBR) to a set of N destinations
selected randomly at the beginning of the simulation.
The mobile agent started at a random location and it
tried to migrate to an alternate host after any packet
transmission, but in not less than τ
m
seconds. Pa-
rameter τ
m
defines the minimum interval between mi-
grations (τ
−1
m
is the maximum migration rate). Note
that it is not guaranteed that a migration will occur
every τ
m
seconds. A migration will occur only if a
better location can be recognized with the informa-
tion available from the CPN. Individual transmissions
used unicast and they were sent one after another to
the members of the group each time.
0 0.05 0.1 0.15 0.2 0.25
10
15
20
25
30
35
40
45
maximum migration rate
group message distance (hop count)
group size: 2
group size: 4
group size: 8
Figure 1: Average message group distance (hops) versus
τ
m
.
0 0.05 0.1 0.15 0.2 0.25
2
4
6
8
10
12
14
16
18
20
22
maximum migration rate
average migrations
group size: 2
group size: 4
group size: 8
Figure 2: Average number of migrations versus τ
m
.
The simulation allowed to store the last 20 routes
discovered by smart packets. The stored routes were
accessible to the mobile agent to build a partial view
of the network to select an appropriatehost to migrate.
The experiments considered mobile agents that used
migration to try to reduce the group distance calcu-
lated from the individual path delays to each of group
members. The group size was 3 with τ
m
= 2sec.
Each message sent by the mobile agent to the
group was tagged with an identifier and creation
timestamp to allow the calculation of the group dis-
tance from the individual end-to-end delay observa-
tions. The average group distances that were observed
during the simulations are depicted in Figure 1. The
group distance tend to decrease with increasing values
of the maximum migration rate.
The average number of migrations that occurred
for a chosen value of τ
−1
m
is shown in Figure 2. As
expected, it was observed a clear increase in the num-
ber of migrations with larger values of the maximum
migration rate. It was harder for the algorithm to find
suitable migrations with larger groups as evidenced
by the higher number of migrations observedfor those
groups. On the other hand, there was a small decrease
in message throughput with larger migrations rates as
a consequence of the migrations cost Figure 3.
IMPROVING THE LOCATION OF MOBILE AGENTS ON SELF-AWARE NETWORKS
241

0 0.05 0.1 0.15 0.2 0.25
2.4
2.5
2.6
2.7
2.8
2.9
3
3.1
x 10
4
maximum migration rate
individual message throughput (bps)
group size: 2
group size: 4
group size: 8
Figure 3: Message throughput (bps) versus τ
m
.
4 CONCLUSIONS
As the volume of business traffic on the Internet con-
tinues to increase with a shift from centralised to dis-
tributed architectures, better solutions are needed to
handle group communications under uneven network
conditions. The paper has explored the use of a two-
tier system to support QoS-critical group communi-
cations that often arise in distributed systems.
The use of a CPN-based mesh offers self-aware
communications with adaptive routes that change
over time and that attempt to take advantage of re-
sources that may become temporarily available in the
network. The paper suggested a way to exploit the
paths dicovered by CPN to determine appropriate mi-
gration hosts for mobile agents. The approach consid-
ered the progressive improvement of the location of
mobile agents by making selective migrations, which
are calculated from the focalised network information
collected by CPN or comparable self-aware network.
Finally, a comprehensivesimulation study showed
the advantages of this approach in the context of im-
proving a critical group communication.
ACKNOWLEDGEMENTS
The work presented in this paper was partially sup-
ported by the project CASCADAS (IST-027807)
funded by the FET Program of the European Com-
mission.
REFERENCES
Gelenbe, E., Lent, R., and Xu, Z. (2001). Measurement and
performance of cognitive packet networks. J. Com-
puter Networks, 37:691–701.
Gelenbe, E., Lent, R., and Xu, Z. (2002). Performance and
measurements of cognitive packet networks. Com-
puter Networks.
Gonzalez-Valenzuela, S. and Vuong, S. T. (2002). Evalu-
ation of migration strategies for mobile agents in net-
work routing. In MATA ’02: Proceedings of the 4th In-
ternational Workshop on Mobile Agents for Telecom-
munication Applications, pages 141–150, London,
UK. Springer-Verlag.
Gray, R. S. (1995). Agent Tcl: A transportable agent sys-
tem. In Proceedings of the CIKM Workshop on Intel-
ligent Information Agents, Fourth International Con-
ference on Information and Knowledge Management
(CIKM 95), Baltimore, Maryland.
Gray, R. S., Cybenko, G., Kotz, D., Peterson, R. A., and
Rus, D. (2002). D’Agents: Applications and perfor-
mance of a mobile-agent system. Software— Practice
and Experience, 32(6):543–573.
http://www.JumpingBeans.com, O. Jumping beans.
http://www.recursionsw.com/, O. Voyager.
Lange, D. B. and Oshima, M. (1999). Seven good reasons
for mobile agents. Commun. ACM, 42(3):88–89.
Lent, R. (2008). The integrated network-environment sim-
ulator (ines), online: http://san.ee.ic.ac.uk/ines.
Peine, H. (2002). Application and programming experience
with the ara mobile agent system. Softw. Pract. Exper.,
32(6):515–541.
Ranganathan, M., Acharya, A., Sharma, S. D., and Saltz,
J. (1999). Network-aware mobile programs. pages
567–581.
Sudmann, N. P. (1996). TACOMA - fundamental abstrac-
tions supporting agent computing in a distributed en-
vironment. PhD thesis.
Tai, H. and Kosaka, K. (1999). The aglets project. Commun.
ACM, 42(3):100–101.
Wong, D., Paciorek, N., Walsh, T., DiCelie, J., Young, M.,
and Peet, B. (1997). Concordia: An infrastructure for
collaborating mobile agents. In MA ’97: Proceedings
of the First International Workshop on Mobile Agents,
pages 86–97, London, UK. Springer-Verlag.
ICEIS 2008 - International Conference on Enterprise Information Systems
242