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

Re: conceptual questions



Kedar,

first of all, thanks for the prompt reply.

> > (f) ...in case of a host receiving packets, addressed to a local
> >        agent, a transport layer sink. Is it possible that
> >        packets (from multiple links) arrive at a host at
> >        a rate which is superior to the rate in which the
> >        sink agent is receiving such packets? Were drops modeled,
> >        in which queues that would be?
> >
> The sinks can absorb the packets at any rate and hence the aggregate of
> all the incoming link rates. There are no buffers within the node and
> no drops are modeled within the nodes right now.

Ok. Suppose then that it's not a sink, but instead an agent of another
transport-level protocol. And that this agent receives packets from 
other agents.
Would ns allow me to set an upper-bound rate in the capacity of the 
transport agent at this host, so that i could simulate implosion,
or i'd simulate this within the code of the transport agent itself?

For example, how one would add the simulation of implosion losses to 
SRM? In the code of SRM or ns? any ideas?


> > (g) ...all buffer management is done in a packet-basis, not byte,
> >        so that the packet size would not matter in regards to queue sizes
> >        and drops?
> >
> If you use a Drop Tail (FIFO) queue that is true. But if you use RED you have
> an option of managing the buffer in Bytes. So the buffer management is specific
> to the queue that you are using.

how could i relate this to current ip routers? in a discussion with someone 
else about simulations and routers, i am told...

> 1/ a single shared buffer pool is to osimple - lots of router
> implementations have a n-tier pool of common fixed lenght buffers -
> e.g. minimum, average and max MTU buffers (for TCP ack, 40byte+mac
> header, for average - say 576+ headers, and for max, depends on
> interfaces.....ATM might be 9128, ether 1546 etc)
>
> 2/ can do congestion control on input queues in SOME newer routers
> 
> 3/the time to get a packet between input queue and output queue MAY be
> varialbe with long tai l(e.g .shared bus router, or router where the
> bus is preemtpted for the CPU to get at memory for a long computation
> such as dijkstra SPF, as in some cisco models....)


could you please comment that with ns in mind?

many thanks,
marinho

-- 
==== Antonio Marinho Pilla Barcellos ==================================
Ph.D. Student                                   phone: +44 191 222 8007
Department of Computing Science                 fax:   +44 191 222 8232
University of Newcastle upon Tyne   WWW: http://www.ncl.ac.uk/~n4521529
The United Kingdom NE1 7RU      e-mail: [email protected]  
=======================================================================