Databases Reference
In-Depth Information
Within each host, the HGN subnet is structured as an executable code, and
each neuron is represented as an associative data structure in a block of mem-
ory space for storing and recalling patterns. The communications between
neurons is achieved using either a sequential or parallel processing approach,
via a message-passing infrastructure, such as Message Passing Interface (MPI).
Each neuron can also be represented as a processing unit in a multi-core pro-
cessor machine. Different configurations of this distributed HGN scheme will
be discussed in the next chapter.
To obtain an overall view of patterns, the distributed HGN scheme allows
communications between HGN subnets residing on different hosts. The com-
munications involve message exchanges containing indices obtained from each
of the HGN subnets for every subpattern analyzed by the network. Cumula-
tively, these indices represent the entire pattern structure.
The accuracy and scalability of a distributed approach for the HGN algo-
rithm have been verified using two significant factors related to the deployment
of an application on any distributed systems:
1. The varying capabilities of the participating nodes
2. The distribution of the computational load
Two distributed schemes were simulated. The first verification addresses vary-
ing processing capabilities within a distributed system through the non-
uniform approach. The second demonstrates the distributiveness of the ap-
proach through the uniform distributed HGN model. The following section
describes the design of this distributed HGN approach by considering these
two factors.
4.4 Design Evaluation for Distributed DPR Approach
We consider two different network distribution approaches for distributing
the HGN network into different subnets. These approaches are the uniform
and non-uniform approaches.
4.4.1 Non-Uniform Distribution
The distributed approach in HGN takes the form of multiple HGN com-
positions. These compositions are distributed across the network. In the non-
uniform model, the compositions may vary in size. For this simulation, a
7-21-7 composition was chosen; there are three sub-structures, comprising
two 7-element HGNs, and one 21-element HGN. Figure 4.9 illustrates these
compositions. Note that in this diagram, the middle host/network has been
determined to be substantially larger than the other two hosts/networks.
