ARPANE Review. Di Niu

This paper first points out the shortcomings of SPF, an ARPANET
routing algorithm based on periodic measurements of link delays. Such
shortcomings are primarily concerned with highly loaded traffic.
According to the authors, when traffic load is high, PSNs tend to
have homogeneous behaviors for SPF. Thus, there will be so-called
"routing oscillations", which lead to the over-utilization and under-
utilization of some links. Moreover, a bigger amount of control
overhead will be incurred due to the need of frequently satisfying
"update generation threshold criterion". It then proposes a
modification of SPF by revising the link delay reporting scheme,
while keeping the original methods of computing routes in SPF. The
main idea is that in order to subdue "routing oscillations", the new
scheme limits the changes between successive reports of link delays
so that these changes are always moderate. Simulation results are
offered to evaluate the performance of the new method.

However, in my personal view, this paper is not of high quality at
all because of several reasons. First of all, the paper is too
verbose at some point, while at some other points, it fails to
provide concrete supporting facts. For example, throughout Section 3
and 4, the paper keeps talking about a simple fact of routing
oscillation in the highly loaded case, whereas it does not provide
any support for its argument concerning control overhead or CPU
utilization of SPF. Such paragraphs as the 3rd paragraph of section 4
is not clear at all.

Second, the new scheme may not be a good scheme at all. Apparently,
it will perform very badly in the cases of lightly or moderately
loaded traffic because of the limits on the change between successive
reports. No analysis is provided to show whether the new scheme is
optimal in the heavily loaded case either (actually it can not
guarantee optimal routing at all.) Because so much meaningless and
onerous calculation is involved, it is very likely that the scheme
may only suitable to specially contrived network parameters.

Moreover, the new scheme requires a central agent to collect reports,
to compute some function of them and to disseminate the computed
information through a flooding system. This could incur much more
delay in a big network, and is not scalable at all, not to mention
the computational overhead it introduces.

In order to improve the paper, the authors need at least provide
either measurement based evidence or mathematical analysis to support
their arguments. However, this may still not be enough to compensate
a poor idea.

Di Niu
Received on Tue Sep 19 2006 - 01:26:34 EDT