Summary: | 碩士 === 國立台灣工業技術學院 === 電子工程技術研究所 === 85 === A mesh architecture is one of popular parallel topologies.
Because it hasnot only simple and regular architecture but also
better scalability than ahypercube architecture does, the mesh
has become one of the most populararchitectures for massively
parallel computers (MPCs). Multicast is a
collectivecommunication service that the same message is
delivered from a source nodeto an arbitrary number of
destination nodes. There are unbalanced trafficpatterns under
prior multicast strategies in 2D mesh multicomputers. Either
only a shortest path can be chosen or some channels have heavy
traffic load when the progress ofthe messages are delivered.
This situation will lead to starvation,moreover,the messages may
cause deadlocks (such as tree-based multicast),leading to
reducing the system performance, severely. In this paper, we
present a balanced and adaptive multicast wormhole routing
strategy in 2D mesh multicomputers, called Hamilton-based two-
phase multicast (H-TPM). This strategy uses Hamiltonian paths
and wormhole routing mechanism, and can send messages to
arbitrary number of destination nodes within two start-up
communication phases without any virtual channel. In order to
balance multicast traffic, H-TPM allows some intermediate nodes
which are not in the destination set, called the dominating
nodes, to do some operations. Therefore, the system performance
will be improved significantly. We evaluate the performance of
our strategy with respect to prior ones, and the simulation
results indicate that ours exhibits better performance than the
others.
|