[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: active networking in ns



[ Sorry for the noise, there is no need for forwarding messages, I found the start
of the thread which I had missed among 'thousands' of other messages; I guess the
rest of my mail remains valid though ]

Lidia Yamamoto wrote:

> Guillermo,
>
> Could you please send me, or post to the list in case you judge more people might
> be interested, the initial messages that lead to this thread? (the first message
> I got is yours, which is already a reply to a reply).
>
> Are you talking about PANAMA's AN package for ns (Diane Kiwior)? I also studied
> this implementation, but it looks quite specific to reliable multicast, and I
> found it difficult to apply to a more generic context.
>
> Therefore I started working on a generic implementation of Active Networking in
> ns, which is intended to be as simple&stupid as possible, while fitting the goals
> of enabling generic active network packets to be transmitted and executed, using
> the IP-like packet forwarding model of ns. My simulation does not require
> modifying the classifier for the moment, although I guess that for optimisation
> purposes I'll probably end up by doing it.
>
> What I do is to define "execution environment" agents (implemented as a subclass
> of Agent/Message) that send "capsules" among themselves. These "capsules" are
> simply Agent/Message packets in which the "message" is the code to be executed.
> For the moment the code I send is only the name of a tcl procedure followed by
> parameters (therefore code loading/caching is not simulated). Using this it is
> surprisingly simple to implement typical AN examples such as ping, multicast,
> etc.
>
> For the moment I solve the routing problem by having the execution environments
> "cheat": they look at the global ns static routing tables to find the next "EE
> hop" on the path towards a destination EE. A more elegant solution would be to
> deploy an active routing protocol using IP as the link layer, but since I don't
> need sophisticated routing at the moment, I can live with the "cheating"
> solution. The drawback of this solution is that I cannot benefit from IP
> multicast, so the multicast trees are not optimal, the same as when using MBONE
> tunnels.
>
> If you and/or other people are interested I am glad to make my implementation
> available, but in this case I need some time to prepare a neat version for
> distribution...
>
> Lidia.
>
> Guillermo Rodriguez Garcia wrote:
>
> > >> I have been studying this for a while in the past days, and there
> > >> are only two things about it:
> > >>
> > >> - Although they intended to release this package as a 'generic'
> > >> extension to ns, in fact it is very focused towards AER, their
> > >> own active protocol; or at least towards active protocols for
> > >> multicasting.
> > >
> > >The basic Active Network Package is designed to allow multicasting
> > >in active networks so there is some additional overhead in that respect.
> >
> > Yes; I guess waht I will have to do at the end is just use
> > your AN extension package but I will remove everything related
> > to mcast to keep it as simple as I can.
> >
> > >> - Also, this solution involves a modified classifier.cc, which
> > >> I would like to avoid unless I'm forced to use it (same reason
> > >> as above).
> > >
> > >I don't think there is any way around this without modifying classifier.
> > Since
> > >all packets released by an agent in a node pass back through that node's
> > entry
> > >point, for active networking the entry point must be modified.   If a
> > packet is
> > >an active packet that has not been seen by the active agent at that node, it
> > >should be forwarded to the active agent.  If a packet is an active packet
> > that
> > >has been seen by that node's active agent, it should be forwarded to the next
> > >hop.  We added two fields to packets and modified classifier.cc to allow the
> > >entry point to make that decision.
> >
> > Someone suggested here to attach the agent to the queue, instead of
> > attaching it to the node. It looked like a good idea, but I am just
> > realizing now that this approach would need a much higher number of
> > agents (two per link instead of one per node), which wouldn't be a
> > good thing for a large scale simulation :-(
> >
> > So I'm afraid that you are right and I cannot avoid having a modified
> > classifier.cc.
> >
> > BTW, are you still working with active networks ? I have been
> > studying you ants-based AER protocol and there are two or
> > three questions that I would like to ask you, if you have the
> > time.
> >
> > Thanks for your reply !
> >
> > Regards,
> > G.
> > --
> > Guillermo Rodriguez Garcia
> > <[email protected]>
> >
> > --
> > Guillermo Rodriguez Garcia
> > <[email protected]>