Hi Yongli,
[speaking as an individual and co-author]
The Parent PCE can determine a likely domain path based on
its knowledge of the inter-domain links. However, the
domain path may not be optimal because it does not know
the detailed information in each domain.
I think there is something missing here. There would be zero value in
this
hierarchical PCE approach if it could not generate optimal paths!
The parent PCE is able to determine all potential end-to-end domain
paths.
It can then ask the child PCEs responsible for each domain to compute
optimal paths across the domains and return those paths together with
their
costs.
The parent PCE can combining these paths to construct a set of potential
end-to-end paths and can easily determine which is optimal by summing the
costs of the segments.
Of course, the parent PCE does not need to perform this work for *all*
potential domain paths if it can apply policy to guess that some domain
paths are likely to result in suboptimal solutions.
Furthermore, the parent PCE can make a trade-off. By issuing computation
requests to the children one at a time and using the returned information
to
prune domains from the set of potential domain paths. This (obviously)
reduces the number of child PCEs involved and simplifies the computations
they have to perform, but may take longer to reach a solution.
Maybe there is a balance between the optimization and
the complexity. Do you think so?
This is always the case. But we aim to produce a simple (non-complex)
architecture that can be used in complex or simple environments, and that
can provide optimal solutions.
I am always wondering about the optimal end-to-end path in
the current framework.
Another question. Can we abstract the detailed information
in each domain into some simple forms and provide them to
the parent PCE? For example, if we compute the shortest
path according to the hop numbers. We can gain the hop
numbers between each border nodes pair in each domain by
the child PCE, and provide them to the parent PCE. Then
the parent PCE can gain the optimal the end-to-end path
across multi-domains.
I am strongly opposed to this idea. Operating on a single metric like hop
count, it appears that this approach might be practicable. It might
suffer
some instability during link failures, but this could be damped leading
to
sub-optimal paths in those windows.
But what value is this information is a sophisticated TE environment? If
your definition of an "optimal" end-to-end path is limited to just least
number of end-to-end hops, this mechanism can work. But the constraints
on
TE paths include *at*least* bandwidth availability as well. It becomes
radically more complex to abstract the available bandwidth and tie it to
the
hop counts for all edge-to-edge connectivity across a domain. It is also
likely that this information will change every time some unrelated LSP is
set up within the domain. The result is a very large data set with a high
rate of change and non-trivial abstraction processing.
Of course, an option is to aggregate the information by introducing
approximations. This can reduce the size of the data set considerably,
and
will also reduce the flapping, but might not reduce the processing
requirements. BUT, the main concern with such aggregation is that it
loses
information; and any loss of information risks the optimality of the
final
solution.
Moreover, when we extend the PCE model beyond "simple" packet TE
environments to use cases with multiple constraints (consider the WSON
work
currently in hand) this abstraction becomes hopelessly complex.
So, the solution we have developed uses just a topological abstraction
(i.e., what inter-domain links exist). It leaves every other item of the
determination of cross-domain paths as the responsibility of the child
PCEs.
Hope this helps.
Cheers,
Adrian
