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

[ns] Re: Multiple Multicast sessions




i do not use CBR, but my own traffic generator which I call VBR.
it generates packets with an average rate A. 
Specifically, it uses a parameter P (I use P=3, or P=6),
and generates 1 packet with probability 1/P and P*A+1-P packets with
probability 1-1/P.

so if we have P=3, and i want 32Kbps avg rate, the source can generate
upto 10 packets at the SAME simulated time for the same source.

in addition, i have upto 48 such sources, with avg rates ranging from
32Kbps to 1024 Kbps.

so what happens is that the session with the smaller id, gets to send
packets first, so the number of packet events generated by a single source
at the SAME simulated time can fill up the whole queue.

in real life what we may expect is that "different" sources get to this
chance at different times. in this case what happens is that the same
source gets the chance EVERY time because of the way things are scheduled.
and so i get a skewed performance.

i may have to make changes in the scheduler, but i am apprehensive about
touching a core component of ns.

-srinivasan


 On Mon, 12 Jun 2000, Haobo Yu wrote:

> > what i find is that sessions with "smaller" id (i.e., those which were
> > started first by the simulator based on the for-loop variable) have lower
> > losses.
> 
> I guess you must be using CBR sources right? In such cases because you do
> not insert any randomness into packet generation the ordering of packets
> on the aggregate link from all flows are deterministic and you should
> expect to see the same drop pattern throughout your simulation. This would
> result in what you've seen.
> 
> Now you are in charge of your simulation and you can change your cbr
> traffic source to add in randomness, say, packet intervals, while maintain
> the same average rate. Is that one of the reasons you saw randomness in
> real-life traffic?
> 
> Please let me know if I missed anything.
> 
> - Haobo
> 
> > The obvious conclusion is that the packet events for the sessions which
> > were started earlier are scheduled earlier. since I am modelling bursty
> > traffic, these packets fillup the queue causing other packet events
> > scheduled at the same simulated time, to incur losses.
> > 
> > In a real-life situation one may expect these losses to be distributed
> > "randomly" among all the competing sessions. so how do i make this random,
> > instead of this deterministic outcome?
> > 
> > I tried using a "Heap" scheduler, but the results are the same.
> > is there anyway out of this?
> > Any help will be appreciated. Thanks in advance.
> > 
> > -Srinivasan
> > 
> > 
> 
>