Database Reference
In-Depth Information
A structured P2P network, as shown in Fig.
8.1
c, uses a DHT to link nodes so
that a query can be effectively and efficiently resolved. This will be discussed in the
following section.
8.3
Adaptive Image Retrieval in a Self-organizing
Chord P2P Network
8.3.1
System Architecture
In a distributed system, images are located at participating nodes across the network.
It is impractical for an image query to traverse all the participating nodes due to
the high communication cost. Thus, one way to make a distributed search system
efficient is by performing a pre-computation, a
search index
before queries are made.
While answering quires, these indices can be used to precisely locate the nodes
serving the matching images without having to communicate with all the nodes.
Instead of storing the entire index in one server, a distributed indexing system
distributes the index to multiple nodes in the network. Thus, the challenge is finding
the right scheme to partition the index across the nodes in the network. A solution is
to partition indexes by
cluster identifications
(IDs) of an image and store all index
entries of a given cluster ID on a specific node. In this scheme, a user searches for
images using the cluster ID of a query image, while the system utilizes the cluster
ID as a
key
. To process a query, the system needs to fetch the image list for the key
from the network. Once the list has arrived, the images that appear in the list are the
matching images for the query.
This section presents the Cluster-identification Search System (CSS), that gen-
erates index entries with
cluster ID
as a key. Figure
8.2
shows the overview of
the CSS, which proceeds in two stages, indexing and query processing. Due to
its simplicity and elegant approach to routing queries, the Chord protocol [
226
]
is selected to implement the system. The processing is conducted between the
application layer and the DHT layer. In the indexing stage in Fig.
8.2
a, each node has
a local image database to be shared with other nodes. The image feature extractor
accesses each image in the local database and then performs mapping via SOTM
(discussed in Chap.
3
). This results in a cluster ID of the best matched SOTM cluster.
The mapping provides an index entry
for the input image, where
clusID
is the cluster ID, and the URL contains IP addresses, port numbers, and
image names of the owner node. It also applies to all images in the local database,
resulting in index entries corresponding to various cluster IDs. All index entries
for a given cluster ID will be stored on a particular node. The node is chosen
by hashing the cluster ID, using a hashing function and Chord. Specifically, the
command
put(key, data)
is the application programing interface (API) that
lets the application layer put a data item (i.e., index entry) in the nodes in the P2P
network. DHash layer works by associating the keys with data items in the nodes.
<
clusID
,URL
>
Search WWH ::
Custom Search