DISTRIBUTED ALLOCATION OF A CORPORATE SEMANTIC
WEB
Ana B. Rios-Alvarado, Ricardo Marcelín-Jiménez and R. Carolina Medina-Ramírez
Department of Electrical Engineering, Universidad Autónoma Metropolitana-Iztapalapa, Atlixco186, DF, Mexico
Keywords: Semantic Web, Ontologies, Distributed Storage.
Abstract: This paper outlines a general method for allocating a document collection over a distributed storage system.
Documents are organized to make up a corporate semantic web featured by a graph
1
G , each of whose
nodes represents a set of documents having a common range of semantic indices. There exists a second
graph
2
G
modelling the distributed storage system. Our approach consists of embedding
1
G into
2
G , under
space restrictions. We use a meta-heuristic called “Ant Colony Optimization”, to solve the corresponding
instances of the graph embedding problem, which is known to be a NP problem. Our solution provides an
efficient mechanism for information storage and retrieval.
1 INTRODUCTION
The semantic web is an extension of the current web
intended to provide an improved cooperation
between humans and machines (Berners-Lee, 2001).
This approach relies on a formal framework where
information is given a well-defined meaning. It uses
ontologies to support information exchange and
search. It is also based on semantic annotations to
code content representation. In addition, it works
with formal knowledge representation languages in
order to describe these very ontologies and
annotations.
There exist diverse contexts that may profit from
the ideas developed by the semantic web. Corporate
memories, for instance, share common features with
the web, both gather heterogeneous resources and
distributed information and have a common concern
about the relevance of information retrieval.
Nevertheless, corporate memories have an
infrastructure and scope limited to the organization
where they are applied. Among the heterogeneous
resources belonging to a scientific group or
enterprise, for example, documents represent a
significant source of collective expertise, requiring
an efficient management including storage,
handling, querying and propagation. From this
meeting between the web and the corporate
memories a new solution is born, the Corporate
Semantic Web (CSW). Formally, a CSW is a
collection of resources (either documents or
humans) described using semantic annotations (on
the document contents or the persons
features/competences), which rely on a given
ontology (Gandon, 2002).
This paper describes a general method to allocate
a CSW in a distributed storage system. A system
such as this is a collection of interconnected storage
devices that contribute with their individual
capacities to create an extended system offering
improved features. The importance of this emerging
technology has been underlined in recent research
works. Although its simplest function is to spread a
collection of files across the storage devices attached
to a network, desirable attributes of quality must also
be incorporated.
Our proposal, that we call a semantic layer, is
mainly defined by its two main procedures:
information location and query. Location solves
document placement and creates the tables
supporting query. We consider this approach
provides a flexible, scalable and fault-tolerant
service. Location is solved using a meta-heuristic
that accepts either a centralized or distributed
implementation. We believe this may lead to a self-
configurable semantic storage system.
The remaining of this paper is organized as
follows. Section 2 is an overview of the previous
work which is related to our proposal. Section 3
presents the general structure of our semantic layer.
Section 4 is a formal statement of the Graph
Embedding problem. Section 5 is about our
simulation method. Section 6 describes the
173
Rios-Alvarado A., Marcelín-Jiménez R. and Medina-Ramírez R. (2009).
DISTRIBUTED ALLOCATION OF A CORPORATE SEMANTIC WEB.
In Proceedings of the International Conference on Knowledge Management and Information Sharing, pages 173-179
DOI: 10.5220/0002302501730179
Copyright
c
 SciTePress

experimental agenda and results. Section 7 closes
our paper with a summary of our contribution and
directions for further work.
2 RELATED WORK
Semantic storage on P2P systems is a very active
area of research. Different proposals have been
developed in order to address the many aspects of
this issue. Risson and Moors focused on information
retrieval (Risson and Moors, 2006), based on the
utilisation of semantic indices. They devised two
alternative mechanisms called VSM (Vector Space
Model) and LSI (Latent Semantic Index). In VSM,
any document or query is featured by a vector of
terms or indices. Therefore, a query defines a point
in the corresponding vector space. In contrast, LSI
performs a keyword query not only on the indices
that feature the documents, but also on the contents.
This way it is possible to find documents having a
semantic proximity with the query, although they
may not have any index matching the initial quest.
PeerSearch (Tang, 2002) is a project that included
both mechanisms to support document indexing.
In (Wolf-Tilo, 2005) we found an information
recovery service based on the classification of
contents in a digital library, on a P2P system too.
There exist projects like (Kjetil, 2006) and (Crespo
and Garcia-Molina, 2002) advocating the utilization
of semantic overlays on top of P2P storage
networks, already deployed. In SOWES (Kjetil,
2006), for instance, peers are gathered in
neighbourhoods or zones, based on keyword vectors
describing their contents. Next, each zone collects
vectors from its corresponding nodes and rebuilds
new clusters based on vector similarities. Also,
(Crespo and Garcia-Molina, 2002) propose a
semantic overlay network (SON) on top of a music
storage system. This approach allows queries to be
forwarded to the corresponding SON. A SON
creation is based on the hierarchical classification of
those concepts describing the semantic profile of
potential documents. Then, files are stored under
their corresponding SON. Retrieval starts with a user
query which is classified, i.e. linked with a key
concept within the hierarchy.
Heterogeneity is another important matter.
Although there might be many P2P storage systems
supporting semantic retrieval, it does not mean that
exchange is possible between any two of them.
Besides its own resources and the description of the
underlying knowledge domain, each local system
must adopt a “trade” convention in order to be a part
of an exchange agreement. In a Peer Data
Management Systems (PDMS), for instance, each
member has a mapping mechanism that translates
any external query into its local scheme. With this
approach, the extended system is able to render
documents to any query, from any reachable place.
Piazza (Halevy, 2003), Edutella (Nejdl, 2002), and
RDFPeers (Cai, 2004) are extended systems, built on
these principles.
Finally, ANTHILL (Montresor, 2001), is a
framework providing support for P2P applications
development. It is based on a biological model: ants
foraging behaviour. It considers low level features
such as communications, security and scheduling.
As for document storage, ant algorithms follow a
draconian policy where those files frequently used,
are preferred over those seldom consulted.
3 SYSTEM DESCRIPTION
It is a well-known fact that the IP routing task, at the
Internet, is supported by two complementary
procedures: the first one, which performs table
keeping and the second one, which performs table
querying. Following similar principles, we propose
the storage of a given CSW based on two
procedures. First, we solve document location and
build a table, whose entries show the places in
charge of a given set of documents. Second, we
perform look-up on this table in order to consult the
corresponding contents.
This paper addresses our methodology for
document location. From our view, a given CSW is
a set of documents classified according to an
underlying ontology. This ontology itself describes a
collection of concepts (categories) which can be
mapped to an ordered set of indices. Each document
is therefore labelled with an index. Let us assume
that the CSW is partitioned in such a way that, all
the documents sharing a common index are
associated to the same subset. Now, the CSW can be
modelled by means of a graph
1
G , each of whose
nodes represents a subset of the partition. Every
documental node
1
Gv
i
∈
is featured by 2 parameters,
the range
21
...
ii
rr of semantic indices spanning its
corresponding documents and the weight
)(
i
v
ω
,
which is the information these documents amount to.
Figure 1 is an example of a given CSW, once it
has been prepared for storage. There exist 17
different concepts (categories) which, in turn, define
17 indices. Nevertheless, the whole set of documents
it is encompassed by a graph having only 5 nodes,
KMIS 2009 - International Conference on Knowledge Management and Information Sharing
174

each of them featured by its range and weight. Two
nodes sharing a link are assumed to have related
concepts.
Figure 1: CSW modelled by
1
G .
Figure 2: Storage network modelled by
2
G .
In the complementary part of our description, we
model the storage network using a second graph
2
G .
Each node
2
Gv
j
∈ , from now on store, has an
associated capacity
)(
j
v
κ
that features the maximal
amount of information it is able to contain. Fig. 2 is
an example of storage network. Each store shows its
capacity. We say that the storage network has
homogeneous capacity if, for any store
j
v , kv
j
=
)(
κ
.
Document location implies the embedding of
1
G into
2
G . This problem consists of using as few
stores as possible in order to place as many
documental nodes as possible inside these stores, in
such a way that their aggregated weight does not
exceed the corresponding capacity. When the
particular instance of graph embedding is solved,
each store receives a copy of the look-up table. Each
row in this table has two parts; the left entry
indicates a semantic indices range, while the right
entry indicates the store in charge of the documents
in this range. Fig. 3 shows how
1
G has been
embedded into
2
G . Based on this solution we have
built table 1.
Any user looking for some content in the CSW
may contact any store and ask for a given index.
Using its local table, the store will be able to
recognize whether or not it keeps the corresponding
documents. In the first case, it immediately turns in
the documents to its client. Otherwise, it works on
behalf of its client and contact the proper store to
retrieve the documents matching the user’s query.
Figure 3: The embedding of
1
G into
2
G .
Table 1: Look-up table.
Indices
range
Store
1..5 s
1
6..15 s
2
16..17 s
3
The down side of graph embedding is that it is an
NP-complete problem. Nevertheless, there exist a
vast collection of meta-heuristics developed to
tackle this family of problems within bounded
accuracy and reasonable complexities. We decided
to address our instances of graph embedding using
the ant colony optimization method (ACO). It is a
probabilistic technique for solving hard
computational problems. ACO is inspired by the
behaviour of ants in finding efficient paths from
their anthill to the places where they collect their
food. ACO dates back from the pioneering work of
Dorigo (Dorigo, 1992), but it was a few years ago
when scientists started regarding this technique as a
very promising candidate to develop distributed
meta-heuristics, due to its cooperative and inherently
distributed nature.
4 PROBLEM STATEMENT
Let ),(:
111
EVG be a graph representing a CSW and let
),(:
222
EVG be the graph representing a storage
network. The embedding of
1
G into
2
G is a couple
of maps
),(:
ε
ν
S , where
ν
assigns each node from
1
G to a store in
2
G , while
ε
transforms paths in
1
G
to paths in
2
G , upon the following restriction:
There exists a function
RV 
1
:
ω
, called weight.
Also, there is a function
RV 
2
:
κ
, called capacity.
Let
})(,,{
21 jijiij
vvVvVvvN =∈∈= ν
be the set of
nodes from
1
V
mapped to
2
Vv
j
∈
.
DISTRIBUTED ALLOCATION OF A CORPORATE SEMANTIC WEB
175

2
,)()( Vvvv
j
Nv
ji
ji
∈∀≤
∑
∈
κω
∪
j
j
VN
1
=
',
'
jjNN
jj
≠∅
=
∩
This is, the total weight of nodes in
1
V
stored in
2
Vv
j
∈
, must not exceed the corresponding capacity.
Consequently, our optimization problem can be
defined.
Problem (GE). Find an embedding
S of
1
G into
2
G such that, for a given function RSf : that
measures the cost of a given embedding,
S has the
lowest cost
)(Sf .
For the goals of this work, we will assume that
2
G has homogeneous capacity, this means that each
store has the same capacity
k . Also, we will assume
that there is a linear ordering
L
on the stores in
2
G ,
such that
)(
j
vsucc is the successor of
2
Vv
j
∈
,
according to
L
, or null if
j
v is the last element in
L
.
GE is an NP-complete problem (Savage, 1991).
5 ANT COLONY OPTIMIZATION
Even though it is accepted that the initial work in
ACO was due to Dorigo, we will adopt a description
by Gutjhar (Gutjahr, 1999), which is slightly
different, but fits better with our exposition.
Our implementation consists of creating
Z
scout
ants. Every ant is charged to perform a random
depth first search on
1
G . As each ant travels across
the graph, it associates the visited nodes to a given
store
2
Gv
j
∈ . When the aggregated nodes’ weight
exceeds the capacity of the current store, it reassigns
the last node to
)(
j
vsucc and starts this filling
process over again, provided that there are still
nodes to visit. Every ant exhaustively visits
1
G and
reports its solution path to the common nest. We call
this procedure traversal. Depth first search (DFS) is
the building block of this stage. It is implemented
using a well known distributed algorithm (Cidon,
1988) with time and message complexities
)(nO and )(mO , respectively. Where n is the order
of
1
G , and m its size.
In due time, the nest analyzes the cost of each
reported solution. Then, it spreads more or less
pheromone on each path, depending on its quality,
i.e. according to a defined evaluation function; a
good path receives more pheromone than a bad one.
Prizing, as it is also called this stage, is carried out
using propagation of information with feedback
(PIF) (Segall, 1983). It is a distributed algorithm
with time and message complexities
)(DO and )(nO ,
respectively. Where
D and n are the diameter and
the order of
1
G , respectively. As a by-product, it
also builds a spanning tree on
1
G . From this moment
on, our method profits from this spanning tree, in
order to perform successive prizing.
Then, the nest starts the next cycle with
Z
new
scouts that will perform the same procedure:
traversal. Nevertheless, even though it still is a
random search, the prizing phase biases the new
resulting paths. Rounds are repeated for a fixed
number or until the difference between the best
solution of two consecutive rounds does not exceed
a given bound.
6 EXPERIMENTS AND
ANALYSIS
Once we had a running simulator, we designed a set
of experiments in order to evaluate the quality of our
method and its complexity. Let us recall that, as we
work with pseudo-random number generators, each
individual simulation requires a new seed to grant a
new collection of results. So, from now on, when we
describe a single experiment, we mean the same
simulation repeated under 10 different seeds.
From our point of view, we consider that an
instance of the problem is solved when 75% of the
agents follow the same trail, which represents the
best solution, i.e. the least number of stores from
2
G
able to allocate all the nodes from
1
G .
In the first set of experiments we investigated the
initial number of ants
Z
that produces, during the
first round, the highest variability on the resulting
trails. The explanation is that we wanted to bind the
resources required to produce the biggest amount of
independent random trails (potential solutions) on
1
G . For nodes in
2
G , we fixed their storage capacity
c =500. Next, for nodes in
1
G , we designed three
different initial graphs, with 100, 300 and 600 nodes,
respectively. In turn, each of them produced 5 new
graphs, with common order but random weights
uniformly distributed between 0 and 500. Then, we
tested each of the 15 resulting graphs under 7
different values of
Z
, 5, 10, 15, 20, 25, 30 and 35.
This means that we devised 105 experiments.
According to the different instances of the
problem, and the 7 levels of agents that we tried,
KMIS 2009 - International Conference on Knowledge Management and Information Sharing
176

Table 2: Variance for the number of stores in the first round.
100=n 300
=
n 600=n
Ants(
Z
)
μ
2
σ
σ
μ
2
σ
σ
μ
2
σ
σ
5 60.3 1.63 1.27 158.3 1.32 1.14 308.2 1.43 1.19
10 56.4 0.95 0.97 160.1 2.31 1.52 301.3 3.79 1.94
15 58.1 1.23 1.11 158.9 2.55 1.59 308.6 3.65 1.91
20 59.2 0.89 0.94 157.2 2.40 1.55 304.4 3.82 1.95
25 51.4 0.45 0.67 158.3 1.63 1.27 301.2 5.34 2.31
30 55.5 0.56 0.74 154.4 1.20 1.09 305.1 5.45 2.33
35 59.3 0.50 0.70 156.5 1.26 1.12 307.2 5.36 2.32
table 2 shows the mean(
μ
), variance(
2
σ ) and
standard deviation(
σ
) for the number of stores
required to contain a given number of documental
nodes. This table, as well as, fig. 4, suggests that
there is a minimum number of agents producing the
highest variance. The upper bound of this initial
value can be roughly featured by the expression
O(
n ).
Figure 4: Variance
1
G size 100, 300, and 600.
In the second part of our study we investigated
the relationship between the output and the input of
the problem, i.e. the most compact achievable
embedding for a given set of parameters including
the capacity (
c ) of stores in
2
G , as well as, the
weight (
)(
i
v
ω
) and the order (n ) of nodes in
1
G .
Table 3 shows the parameters under testing and their
corresponding levels. Each individual row represents
a different experiment.
A simple analysis will indicate that the least
number of stores has, as lower bound, the
aggregated weight of all documental nodes, divided
by the individual store capacity. In other words,
1
1
)(
Vv
c
v
stores
i
n
i
i
∈∀≥
∑
=
ω
In this particular case, we can approximate the
summation in the above formula, since we now that
)(
i
v
ω
follows a uniform random distribution
between
[
]
ba, . Therefore, for any i , )(
i
v
ω
can be
approximated by its mean
2)( ba +
. Which, in turn,
produces
c
ba
n
stores
)
2
(
+
≥
The sixth column in table 3 shows the lower
bound on the number of stores, for each of the
experiments consider in this part of our study.
Meanwhile, the fifth column shows the mean value
obtained with our simulations. Notice that the
problem does not have solution when
)(
i
v
ω
can be
bigger than
c .
In the third group of experiments we addressed
the influence of the evaporation factor on the
number of rounds. A high evaporation implies a low
correlation between the outcomes of consecutive
rounds and vice versa. In other words, evaporation
features the memory of the previous findings. This
time, we tried two evaporation strategies: In the first
case, we worked with a fixed factor equal to 0.9. In
the second case, we tried with an initial evaporation
equal to 0.9 which was decreased by 0.1 on each
new round. Table 3 shows the parameters under
testing and their corresponding levels. Again, each
individual row represents a different experiment.
Columns 7 and 8 show the corresponding time
complexities for case 1 and 2, respectively. It is
quite clear that the second approach is always better
than the first one. This new strategy means that we
allow a broad initial quest but, as rounds go by, the
long term memory prevails and we stick to the best
finding to accelerate convergence.
For the second strategy, we evaluated the
covariance (
S ) between n and the number of rounds.
The result
S =399 indicates that there is a direct
dependency between the order of
1
G and the time
complexity. In contrast, the covariance between
c
and the number of rounds is
S = -394.41, which
means an inverse dependency between the storage
capacity and the time complexity.
DISTRIBUTED ALLOCATION OF A CORPORATE SEMANTIC WEB
177

Table 3: Number of stores and rounds.
Ants
(
Z
)
n c
)(
i
v
ω
Number stores Ideal number
of stores
Number
rounds
FE=0.9
Number rounds
variable FE
Number rounds by
multiple linear
regression
10 100 100 0-20 11.31 10.0 10.37 6.42 7.4598
0-50 26.15 25.0 12.65 8.03 7.6398
0-100 53.75 50.0 14.63 9.14 7.9398
300 0-20 4.46 3.33 8.16 3.81 6.2798
0-50 9.25 8.33 9.22 4.76 6.4598
0-100 17.85 16.67 10.64 5.72 6.7598
900 0-20 2.06 1.11 6.45 2.23 2.7398
0-50 3.84 2.78 7.32 2.80 2.9198
0-100 6.35 5.56 8.68 3.38 3.2198
17 300 100 0-60 93.42 90.0 18.18 12.23 8.4998
0-150 - - - - -
0-300 - - - - -
300 0-60 32.45 30.0 11.23 6.47 7.3198
0-150 79.26 75.0 13.24 8.10 7.8598
0-300 152.89 150.0 16.36 9.74 8.7598
900 0-60 13.45 10.0 8.42 4.28 3.7798
0-150 27.69 25.0 9.88 5.36 4.3198
0-300 53.64 50.0 11.04 6.42 5.2198
30 900 100 0-180 - - - - -
0-450 - - - - -
0-900 - - - - -
300 0-180 275.18 270.0 14.62 9.63 10.4398
0-450 - - - - -
0-900 - - - - -
900 0-180 93.64 90.0 10.76 6.87 6.8998
0-450 228.59 225.0 14.89 8.6 8.5198
0-900 455.42 450.0 15.98 10.35 11.2198
We assumed there is a linear model that may
describe the time complexity as a function of
c , n
and
)(
i
v ω , then we used the multiple linear
regression model and obtained the following
expression
)
2
(0120.00059.00040.05298.7
ba
cnrounds
+
+−+=
The last column in table 3 shows the predicted
time complexity for the second strategy, according
to this function. We obtained a correlation
coefficient equal to 73% between simulation results
and the predicted values, which we consider
acceptable, for our goals.
7 CONCLUSIONS
We have presented a general methodology that
enables the operation of a Corporate Semantic Web
(CSW) on top of a P2P distributed storage system.
Our semantic-layer proposal is based on two
procedures called location and query. Each peer
working as a store has a look-up table that supports
query. Contents are ordered according to a semantic
index. Then, the look-up table shows the peer in
charge of the contents associated to a given index.
Nevertheless, the main contribution of this work is
the location procedure that assigns the contents of
the CSW to the corresponding store-peers.
A key hypothesis that may cause debate is that
we assume nodes in
2
G , i.e. stores, as static entities
or, at least, with lifetimes sufficiently long to
validate this assumption. In contrast, many empirical
studies on networks' dynamics tend to show that
unless storage is supported by fixed capacities,
cooperative storage is very volatile. Nevertheless,
these very studies consider that high information
availability can be achieved, even in P2P
environments, by means of data replication. Studies
(Rodrigues, 2005) suggest that when devices offer a
long-term stable service, systems might profit from
“conservative” block-coding. In the other side, those
systems where devices offer an intermittent service
should be built on the bases of “aggressive”
replication. Intermediate solutions, with combined
replication and block-coding techniques, are also
suggested in order to facilitate information retrieval
and tracking.
Therefore, we assume a layered architecture and
consider our proposal working on a “semantic” layer
on top of a “replication” layer. From this
perspective, the higher layer works with static logic-
stores supported by a lower layer dealing with
dynamic, best-effort, storage peers.
Location is modelled in terms of the graph
embedding of
1
G into
2
G
. Here,
1
G represents a
CSW and
2
G is the P2P system. Graph embedding
(GE) is an NP-complete problem that we tackled
KMIS 2009 - International Conference on Knowledge Management and Information Sharing
178

using the Ant Colony Optimization heuristics
(ACO). We evaluated the quality and complexities
of our location method. As ACO consists of ants or
agents that explore the solution space and cyclically
improve the results, we found the best number of
agents that produce the most efficient exploration.
Each round consists of two phases, traversal and
prizing. Next, we devised a prizing mechanism that
accelerates convergence. For the instances of GE
here addressed, we were able to propose a model
that predicts the total number of rounds as a linear
function of each of the parameters under study.
Some important issues remain as directions for
further work. For the sake of simplicity we assumed
that each document in the CSW is labelled with a
single index. What should we do with multi-labelled
or highly nested contents? How should we deal with
the CSW growing? Preliminary work suggest that,
multi-labelled documents can fit well in our look-up
table, by means of linking mechanisms
subordinating all the indices of a document to a
couple of main concepts that define the actual
location of the corresponding file. As for the CSW
dynamics, we consider that storage capacities must
be kept in order to foresee a middle-term growing.
In the long term, it might be the case that the whole
CSW partitioning, i.e. its granularity, should be
redefined and a new allocation procedure might be
invoked. It is also possible that whenever a small
subset of related concepts shows a rapid growing on
the size of their documents, the entire collection
might migrate to a new store node.
Distributed storage is driving many R&D efforts.
From the users’ point of view, it may turn into the
basic mechanism able to unleash the potential
benefits of knowledge management. Health sciences,
agriculture, geomatics, are only a few examples of
the many domains that may dramatically improve
their operations with the adoption of this new trend.
REFERENCES
Berners-Lee, T., Hendler, J., Lassila, O., 2001. The
semantic web.
Scientific American.
Cai, M., Frank, M., 2004. RDFPeers: A Scalable
Distributed RDF Repository Based on a Structured
Peer-to-Peer Network.
In Proceedings of the 13th
international conference on the World Wide Web, New
York, USA
.
Cidon, I. 1988. Yet Another Distributed Depth-First-
Search Algorithm.
Inf. Process. Lett. 26(6)
Crespo, A. Garcia-Molina, H. 2002. Semantic Overlay
Networks for P2P Systems.
Technical report, Stanford
University.
Dorigo, M. 1992. Optimization, Learning and Natural
Algorithms.
Ph.D. Thesis, Dept. of Electronics,
Politecnico di Milano. Italy.
Gandon, Fabien. 2002. ONTOLOGY ENGINEERING: A
SURVEY AND A RETURN ON EXPERIENCE,
Report of research INRIA, team ACACIA.
Gutjahr W. 1999. A generalized convergence result for the
graph-based ant system metaheuristic.
Technical
Report 99-09, University of Vienna
.
Halevy A. Y., Ives Z. G., Mork P., and Tatarinov I. 2003.
Piazza: Data management infrastructure for semantic
web applications.
In Proceedings of the Twelfth
International World Wide Web Conference
(WWW’2003), Budapest, Hungary.
Kjetil, N., Christos, D., Michalis, V. 2006. The SOWES
Approach to P2P Web Search Using Semantic
Overlays.
WWW '06: Proceedings of the 15th
international conference on World Wide Web.
Montresor, A. 2001. Anthill: a Framework for the Design
and the Analysis of Peer-to-Peer Systems.
4th
European Research Seminar on Advances in
Distributed Systems
.
Nejdl, W. Wolf,B. Qu, C. Decker, S. Sintek, M. Naeve, A.
Nilsson, M. Palmer, M. Risch, T. 2002. EDUTELLA:
a P2P Networking Infrastructure based on RDF.
In
Proceedings of the 11th International World Wide
Web Conference, Hawaii, USA
.
Tang, C., Z. Xu, and M. Mahalingam. 2002. PeerSearch:
Efficient Information retrieval in Peer-Peer Networks.
Hewlett-Packard Labs: Palo Alto.
Risson, J, Moors, T. 2006. “Survey of research towards
robust peer-to-peer networks: search methods”,
Computer Networks: The International Journal of
Computer and Telecommunications Networking,
Volume 50 , Issue 17, pp 3485-3521
.
Rodrigues R. and Liskov B. 2005. High Availability in
DHT’s: Erasure Coding vs Replication. IPTPS. 226-
239.
Savage J. E. & Wloka M. G.. 1991. MOB a parallel
heuristic for graph embedding,
5th SIAM Conference
on Parallel Processing for Scientific Computing
.
Segall A. 1983. Distributed network protocols.
IEEE
Transaction on Information Theory IT-29(1):23.
Wolf-Tilo Balke, W Nejdl, W Siberski, and U Thaden.
2005. DL meets P2P - Distributed Document Retrieval
based on Classification and Content.
European
Conference on Digital Libraries (ECDL), Vienna,
Austria
.
DISTRIBUTED ALLOCATION OF A CORPORATE SEMANTIC WEB
179