[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [ns] Huge amount of traffic in simulation?
>> >I have some problems with the huge amount of random traffics simulation.
>> >
>> >For a 50 nodes backbone simulation, every node(with agents) needs to
>> >produce a huge amount of random (or possion distri.) traffics to random
>> >destinations. Also, each traffic holding time is also random. Two ways I
>> >can go:
>> >
>> >1. Each traffic corresponds to an agent pair (traffic sitting on), like
>> > "agent-gen.tcl" script, but the problem is a huge amount of agent pairs
>> > have to be created and attached to nodes. It probably exceeds the
>> > limit of node's port number soon. Seems impossible?
>>
>> - if each pair has only a limited life time during the simulation you could
>> create them and destroy them as needed
>
>A problem is how to destroy them. For node, only "Node::detach .." seems
>to detach an agent, but it replaces it with a null-agent. The port classifier
>space is still unreleased.
there should be a destructor for agents, see agent.cc.
>
>I am doing Optical WDM simulation on a backbone. So, each node acts as
>an access node, where lots of traffics are aggregated with their various
lifetime.
>Also, the traffic(not packet) arrival (start time) is a poison distribution.
For
>ns,
>all traffics are created before the simulation starts. Thus, they all have to
be
>scheduled the start and stop time to the simulator.
you could use a random number generator to schedule random start and stop times
>
>> >2. Using one traffic to simulate multiple traffics. Only create one agent
>> > (like UDP) and one traffic (like CBR) for each node, and manage to
control
>> > the traffic to start and stop randomly. I need modify the traffic...
>>
>> - this has the drawback (I think) of not permitting many flows to be
simulated.
>> >
>> >Does anyone has the similar experience? Which way is better? or some
>> >other better one? Any hints will be appreciated.
>>
>> generating lots of trafic is not nearly as hard as finding a way to trace and
>> (post) process the traces efficiently... I would prefer 1) but it really
depends
>> on what you want to evaluate.
>
>At most time, tracing every packet is not necessary. E.g. in my simulation, I
>just want to get the statistics results, such as wavelength assignment, and
>blocking
>prob. Nam's trace graph is just a demo or education tool, not really useful
for
yes, but to get a statistical result of any value you need events to extract the
stats from... trace-all with a good number of flows for a good simulation time
will produce files the order of Gb. I was not even thinking of Nam here...
>real simultation, since it really needs huge data.
>
>Thanks,
>Bo
Tarik