Model poisson packet arrival in wifi

[Goodman Weng]

Hi, I am a new beginner of NS-3.
I want to model poisson packet arrival( i.e. exponential packet interval) in wifi model.
I think the onoff Application is not suitable, because I can't know STA need how many "On" time in retransmitting the packet.
I try the schedule to do new packet arrival.
Does there exist a more convenience way to do the poisson packet arrival in NS-3?
Thanks for any suggestion.

[Tommaso Pecorella]

Hi,

please take your time to think to the difference between packet generation and packet transmission.

In the textbooks there is often a very bad assumption: that the packets are sent as soon as they are generated. This is plainly wrong in real systems, and while ns-3 doesn't model the process delays, it's still more similar to a real node than you may think.

The ns-3 applications can generate packets with a poisson distribution, but the packets will [most probably] not be sent with a Poisson distribution, simply because of buffers, timers, aggregation, retransmission, and so on.

Unfortunately all these things are part of the Wi-Fi standard, and they can't be easily removed, nor they should.

Of course, this means that the packet collision probability can not be described with a mathematical formula, but that's exactly the point of using a simulator. The real question isn't if the theory and some specially-crafted simulations match (they will, and I'll kick your paper), the question is: is the formula close enough to the real data ?

Cheers,

T.

[Sathyanarayanan Chandrasekharan]

Hi,

Some really good points Dr. Pecorella. 

The poisson distribution assumption for packet generation has always baffled me. Do you know of any sources why is it that the traffic generation distribution considered to be poisson?

Also, there is more confusion as in whether the interval between two packets is poisson or the packet duration is. Your thoughts on these will be appreciated.

Thanks,
Sathya

[Tommaso Pecorella]

Hi,

once upon a time... in a galaxy far far away...

Ok, that's the story as far as I know it.

Poisson distribution models a memoryless process, i.e., one where one "event" is not correlated with another one. An "event" may be anything, actually.

If you pick a large enough bunch of people and you set a average number of calls per hour (note: calls, we're talking about plain old telephone calls), the probability that one user influences another one is practically zero.

In other terms, the number of calls in a given period of time is Poisson.

This assumption holds until the number of people is large enough with respect to the number of calls, i.e., when a single user does a call once in a while.

If the number of calls per user is larger, one should model also the call duration.

The call duration has been (historically) modeled as Poisson as well. This is of course a very rough approximation, and it has been used only because the Poisson distribution has very nice mathematical properties, like mean, variance, memoryless, duality with exponential distribution, etc.

If you try to use more realistic distributions, things get messy, and formulas start to be extremely complex and they require numerical solutions.

That's why Poisson is used: laziness.

About the packet duration (their size) and their arrival process (Poisson or anything), things are messy as well.

For web browsing it has been demonstrated that message distribution is Weibull or Pareto (two helluva distributions to work with, they have no variance and no mean in some cases). Messages are split into packets (fixed size usually) and sent over the air. This leads to a batch process arrival, where events (the messages) can have a given distribution and size, but the actual packets are fixed size, and their number is dependent on the message size.

The message inter-arrival size, in turns, depends on the previous message size (heck) and its transmission delay. Why? Because a user ask for a new webpage *after* he/she has received and read the previous one.

Summarizing, something that can not be mathematically solved. Search for the Web generation process papers, you'll find some.

Note: this is for web... ftp is a different model, netflix anther one, etc.

Hope this helps,

T.