Modeling P2P Grid Information Services with Colored Petri Nets - Grid Computing: Infrastructure, Service, and Applications - page 331

satisfactorily under the harsh conditions of high dynamism produced by
grids. In order to build a P2P-based information service that can survive a
highly dynamic environment posed by grids, a solution needs to be:

• Dynamic and robust, accommodating for high rates of churn and

dynamic information updating.
Responsive, being able to discover resources in the least time •

possible.
Scalable, being able to offer information services for a large num- •

ber of computing nodes.

In this section, we introduce PIndex as a scalable solution for information
services. By structuring nodes into independent peer groups that com-
municate via SOAP [11] messages, PIndex is able to carry out distributed
tasks more robustly. In the following sections, we explain in depth PIndex
architecture, and its fast P2P lookup algorithm.

13.3.1 PIndex Architecture

One key feature of PIndex is its capability to group nodes into peer groups
(PGs), while the entire network can be queried through a fast P2P struc-
ture between PGs. Each PG operates independently with its monitored
data limiting the effect of churn to a PG. Grouping nodes into PGs brings
in three major benei ts. First, the cost of updating dynamic state informa-
tion is greatly reduced since the update in a PG does not have to traverse
the whole network. Second, when a node fails, its effects are kept local
with only the nodes in the same PG needing to be informed to update
their information. Finally, peer nodes might join or leave a PG freely, but
a PG is a stable element in the PIndex network allowing searches to be
performed relative to PGs. This removes the dependency on any particu-
lar node for routing.
Figure 13.3 depicts the architecture of PIndex. Each PG has a unique
identii er (hash key) that is used as a common address in PIndex. PIndex
is structured in such a way that each peer node in a PG has all the infor-
mation about other peer nodes within the same PG. Upon receiving a
request, a peer will be able to answer on behalf of the whole PG, while at
the same time localizing monitoring data to a PG such that any changes to
peers need to be updated only within the PG but not the rest of the whole
P2P network to reduce network trafi c.
As every node in PIndex has information about the resources available
within a PG, fault tolerance can be enhanced through replication. In addi-
tion, the query process can also be sped up since only a single node need
be queried per PG. When a peer node fails, all the references made to
it will be removed dynamically. This is done by an index service using
registration renewal messages as heartbeats. In addition, each node must