[e2e] Address Space Distribution

[Eddie Kohler]

> Many thanks to those who pointed out the recent paper on address structure 
> of IP traffic. I don't think they discuss about address space distribution 
> on network topology but only on address space distribution observed on 
> aggregated IP traffic.
> 
> Anyway it is useful for the part where I need realistic generation of 
> traffic demand per aggregated destination address.
> 
> I have a couple of questions after reading the paper.

Hi, I'm one of the authors on this paper and figured I'd respond to Wumin's
questions.

> 1. In figure 4, does the 'flow' means collection of
>    packets with similar IP 5 tuples (i.e, IP source,
>    IP destination, protocol, dest. port, and source
>    port)?

Yes, exactly.

> 2. In their slides, the generated address structure
>    using multifractal does not resemble the real
>    address structure. However they claimed that
>    they manage to capture the essential properties
>    of the address structure. It seems to me that the
>    density of the address aggregate is quite
>    selective. The concentration is higher from
>    128.0.0.0 to 255.255.255.255 and especially
>    in the region between 128.0.0.0 ~ 176.0.0.0.
>    Any comments?

By "essential properties", we mean the metrics that we developed in the
paper, like aggregate counts (How many /X's were nonempty [contained any
packets at all]?) and, particularly, aggregate population distributions
(How many /X's contained at least Y active addresses?). We talk a bit about
why those metrics might be relevant, and they are the simplest metrics we
could come up with; but of course you could invent a different metric on
which the multifractal model would fail.

The regional differences you're seeing are just class-based allocation
boundaries. Class A (0.0.0.0-128.0.0.0) is relatively unpopulated, Classes
B and C are fully populated, and Classes D and up (multicast et al) are
pretty empty. We have recently looked at modeling these class differences
explicitly. For instance, you can use different multifractal spectra for
the different classes and combine the answers.

Eddie