Chang et al. [5] use both channel and switch congestion
to identify path congestion, and achieves higher saturation
throughput. However, they are easy to fall into the local
optimal problem, as illustrated in Fig. 1(b). Assume that the
target flow f(0,3) intends to send packets from source router
R0 to the destination R3, and six contention flows (f1 to
f6) are competing for the potential paths. At router R0, two
candidate paths are considered, i.e., east channel (from R0
to R1) and south channel (from R0 to R2). When using the
routing algorithm proposed in [5] to find different flow paths,
the south channel will be selected, since it has less congestion
than the other. However, when looking into this problem in
a broader view, it may not be the global optimal strategy,because more congestions may exist in path R0→R2→R3
(four contention flows altogether) than in path R0→R1→R3
(only two). This is due to the intrinsic limitation of these kinds
of routing algorithms, which can only make routing decision
based on the local router or the nearest neighbours.