Re: [ns] On-Off traffic source

[Daniel Ratton Figueiredo <ratton@cs.umass.edu>]

Hi,

> 1. why is the case that "There is always at least
> one packet being sent per on period" for NS EXPOO? 
> Actually if the On period is shorter than the interval,
> it won't generate any packet.

No. The current implementation of the Expoo traffic source always
generates a packet when it enters the on period. Look at the function
next_interval() inside the file expoo.cc. There, you'll find the following
piece of code:

....
/* make sure we got at least 1 */
if (rem_ == 0)
        rem_ = 1;
...

Note that rem_, which is the number of packets remaining to be sent in
this on period is alwyas greater than zero.


> 2. So the difference between NS and the "Deterministic"
> models is just by the "round off errors". 

No, the NS Expoo traffic source does not have an exponentially
distributed ON period, due to its implementation. 


> I can't see
> why by using round off we wouldn't have a "well-defined
> behavior". Can you define (or rephrase) this concept 
> more clearly? 

The point is that the behavior of the NS Expoo traffic source is different
from commonly used on-off sources definitions. This can lead to different
performance menasures when comparing NS simulation results with analytical
or simulation results obtained from other simulators. We encountered this
problem when trying to validate some of our models.


> And even if it is
> not well-defined, a simple modification
> like removing the round-off would work to bring it
> back to "well-defined" models.

Yes, a simple modification, such as making the distribution of the ON
period exponential (not rounded off by the number of packets), would bring
the Expoo source to a consistent definition of an OnOff source. We
sugggest such modification on the Tech Report.

Thanks for your comments,
Daniel and Benyuan


> On Tue, 27 Jun 2000, Daniel Ratton Figueiredo wrote:
> 
> > Hello All,
> > 
> > The following problem came to our attention when we tried to match
> > analytical and simulation performance results with the ones obtained using
> > NS. It turned out that the results were not consistent on models using the
> > NS Traffic/Expoo source.
> > 
> > The problem arrises when using the NS Expoo source, which is an On-Off
> > source, in a simulation model and then validating or comparing the
> > performance results obtained with other simulators or analytical models.
> > The implementation of the NS Expoo source is different from most commonly
> > used On-Off sources, and its specification is also not well documented.
> > This confusion can lead to significant discrepancies in the results due to
> > different source behavior. In the Tech Report below, we point out the how
> > the NS Expoo source implementation differs from commom models, and show
> > that this small difference can impact the performance evaluation of
> > models. Finally, we suggest a small modification to the NS Expoo traffic
> > source to comply with a more traditional and well known On-Off source.
> > 
> > Thanks for your attention,
> > Daniel and Benyuan
> > 
> > 
> > The complete Technical Report, incluiding the modified source code,
> > can be found at:
> > ftp://ftp.cs.umass.edu/pub/techrept/techreport/2000/UM-CS-2000-025.ps
> > 
> > Tech Report Title: 
> > On the Specification of NS and other Known On-Off Sources
> > 
> >                            ABSTRACT 
> > 
> > Traffic source is of great importance to networks modeling. Different
> > traffic characteristics can cause other components to behave
> > differently, resulting in different network performance measures. 
> > On-off source model has been widely adopted in networks research to
> > capture the bursty nature of the network traffic; its behavior is
> > precisely defined. In this work, we carefully analyzed the common
> > structure and the behavior of on-off sources and show that the on-off
> > source provided by ns simulator does not correspond to any of the known 
> > on-off models. This inconsistency between on-off sources has a direct 
> > impact on queueing performance measures. Modelers must be aware of this 
> > fact when using the ns simulator to compare or validate results with other 
> > modeling frameworks. We suggest that the ns on-off source be replaced by a
> > consistent deterministic on-off source; the implementation and analysis 
> > were provided. 
> > 
> > 
> > 
> 
> Guo, Liang 
> 
> guol@cs.bu.edu                     Dept. of Comp. Sci., Boston Univ.,
> (617)353-8924 (O)                  111 Cummington St., MCS-217,
> (617)375-9206 (H)                  Boston, MA 02215
>