Information Technology Reference
In-Depth Information
Moreover, they seem likely to converge. Some of the characteristics of P2P
systems, such as unreliable resources and intermittent participation, pro-
vide important experiences for grid computing and will constitute an even
more signii cant share of resources. Consequently, services, particular in
resource discovery, will have to tolerate failures and adapt to dynamic
resource participation. At the same time, grid scalability is increased and
more communities are allowed to participate.
Four components of P2P resource discovery solution in grids are dei ned
in [22], membership protocol, overlay construction, preprocessing, and
request processing, in order to allow the joining and leaving of nodes at
any time and to recognize the unexplored regions. A commonly used
mechanism in P2P systems is adopted for the membership protocol: a
node contacting its member node to allow joining. The joining member
learns about the neighbors from the membership information sent back by
the contacted nodes.
Overlay construction is built by selecting the collaborating nodes from
the membership list based on desirable characteristics. Much research has
been done [23-25] to construct overlay topology to make efi cient broad-
casting with good performance.
Preprocessing is a technique that enhances search performance before
actually executing jobs. Rewiring overlay construction to update topology
is one kind of preprocessing. Dissemination of resource descriptions to
out of the local area is also a preprocessing technique. However, the pre-
processing strategies may not work well in grids since resources and users
change frequently, making the results of preprocessing inaccurate.
Requests are i rst processed by local components. Unless a perfect match
is found, the request will be forwarded to remote components. A request
is dropped where either there is no further forwarding to be pursued or
time to live (TTL) is equal to zero.
Since there are no central controllers, how to efi ciently propagate
requests becomes a bottleneck.
Figure 7.5
shows the network organization
taxonomy of P2P systems. Generally P2P systems can be categorized as
structured or unstructured networks. For unstructured networks, each
request is l ooded to the whole network. However, since the large amount
of query messages generated by l ooding, this kind of approach does not
scale well. Structured systems use distributed hash tables (DHTs) to main-
tain a structured overlay network among peers and use message routing
to forward the query messages. In a system with
N
peers, each peer in
DHTs only needs to maintain an index for
O
(log
N
) peers and can guaran-
tee to resolve a look-up request in
O
(log
N
) time.
The P2P scheduling infrastructure [26] is based on economic schedul-
ing; however, there is no MetaManager to help forward the request:
1. A consumer submits the job request to a local node.
2. Once the request is received, the node estimates whether it can
i nish the job before the deadline. If it meets the requirement, this
Search WWH ::
Custom Search