The objective of this article is to provide a basis for comparing and
contrasting different approaches to creating robust routing protocols
in massively parallel architectures. Communication bottlenecks form a
major impediment to the development and application of massively paral
lel architectures. Multipath networks using virtual channels and adapt
ive routing protocols will be the rule rather than the exception in su
ch architectures. In multipath networks, adaptive routing protocols ca
n dynamically route messages around network trouble spots such as comp
onent failures or communication bottlenecks. Virtual channels provide
a mechanism for dealing with deadlock as well as increasing effective
through-put via higher link utilization. As a result, these techniques
are increasingly important in providing acceptable communication late
ncy and bandwidth utilization in large networks. Hypercube interconnec
tion networks (HINs) form a large class of flexible and powerful inter
connection networks that have been employed or proposed for several co
mmercial and research endeavors in large-scale parallel computing. The
se networks include the well-known binary n-cube and k-ary n-cubes. Th
is article (1) discusses adaptive routing protocols for HINs; (2) intr
oduces and examines such important issues as communication mechanisms,
the use of virtual channels, deadlock/livelock avoidance in HINs, and
the impact of these issues on adaptive routing protocols; (3) delinea
tes a taxonomy of adaptive routing protocols that serves as the basis
for presentation and discussion, and describes representative protocol
s from each class - emphasizing features common to protocols in that c
lass; and (4) presents and discusses simulation results of the impleme
ntation of representative protocols from each class as well as certain
implementation considerations.