Re: Agents in C++ or OTcl?

[Kannan Varadhan <kannan@ISI.EDU>]

> There are two flavors of simulation in ns: writing Agents in C++ as in
> the tcp related simulations and writing Agents in Otcl as in multicast
> related stuff. I personally prefer the former, but I'd like to hear 
> some analysis from the experts, especially on the performance, memory 
> consumption, debuging, etc.

A lot of how you divide the problem in ns appears to
be a matter of taste and style.  There are some general
principles though:

	- In general, you should place objects related to
	per packet activity in C++, simply for reasons of
	speed.  Per-packet processing via OTcl will make for
	*veeerrrrryyy* slow simulations.

	- Things to do with control, and control traffic, 
	such as dynamic routing can be easily done either
	in C++ Or OTcl.  (The DV routing, RTP, RLM, PIM,
	various multicast routing protocols, etc. are
	mostly in OTcl, and are examples of this type
	of programming paradigm).

	- Most people find themselves having a faster
	turnaround time when prototyping in a scripted
	language.  So, (if you are that sort, or otherwise,
	expect to run into that sort of people), place such
	functionality in OTcl that you expect to play around
	with a lot.  In SRM, (the example I am most familiar
	with), data flow and session message processing was
	in C++, but we want to play with the timer algorithms,
	scalable approaches to sending session messages,
	different loss recovery approaches, etc., and so
	these were done in OTcl.

Generally, performance is slow, and memory consumption is high
for OTcl code (I think Polly mentioned this as well); however,
the advantage is rapid prototyping, with the ability to flexibly
explore the protocol design space quickly.

The complexity of having two languages to work with makes
debugging non-trivial as is, but some hacky solutions are
possible.  Check out the ns-debugging web page (off the main
ns web page) for tips on debugging.


Kannan