debug.tex

柯老师网站上找到的

\chapter{Debugging ns}
\label{sec:debug}

\ns is a simulator engine built in C++ and has an OTcl (Object-oriented
Tcl) interface that is used for configuration and commands. Thus in order
to debug \ns we will have to deal with debugging issues involving
both OTcl and C++. This chapter gives debugging tips at Tcl and C++
level and shows how to move to-fro the Tcl and C++ boundaries. It also
briefly covers memory debugging and memory conservation in \ns.


\section{Tcl-level Debugging}
\label{sec:tcldebug}

Ns supports Don Libs' Tcl debugger ( see its Postscript documentation at
http://expect.nist.gov/tcl-debug/tcl-debug.ps.Z and its source code at
http://expect.nist.gov/tcl-debug/tcl-debug.tar.gz ).
Install the program or leave the source code in a directory
parallel to ns-2 and it will be built. Unlike expect, described in the
tcl-debug documentation, we do not support the -D flag. To enter the
debugger, add the lines "debug 1" to your script at the appropriate
location. In order to build ns with the debugging flag turned on,
configure ns with configuration option "--enable-debug"
and incase the Tcl debugger has been installed in a directory not parallel
to ns-2, provide the path with configuration option 
"--with-tcldebug=<give/your/path/to/tcl-debug/library>".

An useful debugging command is \code{$ns_ gen-map} which may be used to 
list all OTcl objects in a raw form. This is useful to correlate the
position and function of an object given its name. The name of the object
is the OTcl handle, usually of the form \code{_o###}. For TclObjects, this
is also available in a C++ debugger, such as gdb, as \code{this->name_}. 


\section{C++-Level Debugging}
\label{sec:cdebug}

Debugging at C++ level can be done using any standard debugger. The
following macro for gdb makes it easier to see what happens in Tcl-based
subroutines: 
\begin{program}
## for Tcl code
define pargvc
set $i=0
while $i < argc
  p argv[$i]
  set $i=$i+1
  end
end
document pargvc
Print out argc argv[i]'s common in Tcl code.
(presumes that argc and argv are defined)
end
\end{program}


\section{Mixing Tcl and C debugging}
\label{sec:mixdebug}

It is a painful reality that when looking at the Tcl code and debugging
Tcl level stuff, one wants to get at the C-level classes, and vice versa.
This is a smallish hint on how one can make that task easier. If you are
running ns through gdb, then the following incantation (shown below) gets
you access to the Tcl 
debugger. Notes on how you can then use this debugger and what you can do
with it are documented earlier in this chapter and at this URL
(http://expect.nist.gov/tcl-debug/tcl-debug.ps.Z).
\begin{program} 
(gdb) run
Starting program: /nfs/prot/kannan/PhD/simulators/ns/ns-2/ns 
...

Breakpoint 1, AddressClassifier::AddressClassifier (this=0x12fbd8)
    at classifier-addr.cc:47
(gdb) p this->name_
$1 = 0x2711e8 "_o73"
(gdb) call Tcl::instance().eval("debug 1")
15: lappend auto_path $dbg_library
dbg15.3> w
*0: application
 15: lappend auto_path /usr/local/lib/dbg
dbg15.4> Simulator info instances
_o1
dbg15.5> _o1 now
0
dbg15.6> # and other fun stuff
dbg15.7> _o73 info class
Classifier/Addr
dbg15.8> _o73 info vars
slots_ shift_ off_ip_ offset_ off_flags_ mask_ off_cmn_
dbg15.9> c
(gdb) w
Ambiguous command "w": while, whatis, where, watch.
(gdb) where
#0  AddressClassifier::AddressClassifier (this=0x12fbd8)
    at classifier-addr.cc:47
#1  0x5c68 in AddressClassifierClass::create (this=0x10d6c8, argc=4, 
    argv=0xefffcdc0) at classifier-addr.cc:63
...
(gdb)
\end{program}

In a like manner, if you have started ns through gdb, then you can always
get gdb's attention by sending an interrupt, usually a \code{<Ctrl-c>}
on
berkeloidrones. However, note that these do tamper with the stack frame, and on occasion,
may (sometimes can (and rarely, does)) screw up the stack so that, you may
not be in a position to resume execution. To its credit, gdb appears to be
smart enough to warn you about such instances when you should tread
softly, and carry a big stick. 


\section{Memory Debugging}
\label{sec:memdebug}

The first thing to do if you run out of memory is to make sure you can use
all the memory on your system. Some systems by default limit the memory
available for individual programs to something less than all available
memory. To relax this, use the limit or ulimit command. These are shell
functions---see the manual page for your shell for details. Limit is for
csh, ulimit is for sh/bash. 

Simulations of large networks can consume a lot of memory. Ns-2.0b17
supports Gray Watson's dmalloc library (see its web documentation at
http://www.letters.com/dmalloc/ and get the source code from
ftp://ftp.letters.com/src/dmalloc/dmalloc.tar.gz ).
To add it, install it on your system or leave its source in
a directory parallel to ns-2 and specify --with-dmalloc when configuring
ns. Then build all components of ns for which you want memory information
with debugging symbols (this should include at least ns-2, possibly tclcl
and otcl and maybe also tcl). 


\subsection{Using dmalloc}
\label{sec:usedmalloc}

In order to use dmalloc do the following:
\begin{itemize}
\item Define an alias 
\begin{verbatim}
for csh: alias dmalloc 'eval `\dmalloc -C \!*`', 
for bash: function dmalloc { eval `command dmalloc -b $*` }
\end{verbatim}
\item Next turn debugging on by typing \code{dmalloc -l logfile low }
\item Run your program (which was configured and built with dmalloc as
described above). 
\item Interpret logfile by running \code{dmalloc_summarize ns <logfile}.
(You need to download \code{dmalloc_summarize} separately.) 
\end{itemize}

On some platforms you may need to link things statically to get dmalloc to
work. On Solaris this is done by linking with these options:
\code{"-Xlinker -B -Xlinker -static {libraries} -Xlinker -B -Xlinker -dynamic -ldl -lX11 -lXext"}.
(You'll need to change Makefile. Thanks to
Haobo Yu and Doug Smith for working this out.) 

We can interpret a sample summary produced from this process on
ns-2/tcl/ex/newmcast/cmcast-100.tcl with an exit statement after the
200'th duplex-link-of-interefaces statement: 

Ns allocates ~6MB of memory. \\
~1MB is due to TclObject::bind \\
~900KB is StringCreate, all in 32-byte chunks \\
~700KB is NewVar, mostly in 24-byte chunks \\
Dmalloc\_summarize must map function names to and from their addresses. It
often can't resolve addresses for shared libraries, so if you see lots of
memory allocated by things beginning with ``ra='', that's what it is. The
best way to avoid this problem is to build ns statically (if not all, then
as much as possible). 

Dmalloc's memory allocation scheme is somewhat expensive, plus there's
bookkeeping costs. Programs linked against dmalloc will consume more
memory than against most standard mallocs. 

Dmalloc can also diagnose other memory errors (duplicate frees, buffer
overruns, etc.). See its documentation for details. 


\subsection{Memory Conservation Tips}
\label{sec:memconserve}

Some tips to saving memory (some of these use examples from the 
cmcast-100.tcl script): 
If you have many links or nodes: 
\begin{description}
\item[Avoid \code{trace-all} :]
\code{$ns trace-all $f} causes trace objects to be pushed on all links. If 
you only want to trace one link, there's no need for this overhead. Saving
is about 14 KB/link. 

\item[Use arrays for sequences of variables :]
Each variable, say \code{n$i} in \code{set n$i [$ns node]}, has a certain
overhead. If a sequence of nodes are created as an array, i.e.
\code{n($i)}, then only one variable is created, consuming much less
memory. Saving is about 40+ Byte/variable. 

\item[Avoid unnecessary variables :]
If an object will not be referred to later on, avoid naming the object.
E.g.
\code{set cmcast(1) [new CtrMcast $ns $n(1) $ctrmcastcomp [list 1 1]]}
would be better if replaced by
\code{new CtrMcast $ns $n(1) $ctrmcastcomp [list 1 1]}.
Saving is about 80 Byte/variable. 

\item[Run on top of FreeBSD :]
malloc() overhead on FreeBSD is less than on some other systems. We will
eventually port that allocator to other platofrms. 

\item[Dynamic binding (NEW) :]
Using bind() in C++ consumes memory for each object you create. This
approach can be very expensive if you create many identical objects.
Changing \code{bind()} to \code{delay_bind()} changes this memory
requirement to per-class. See \ns/object.cc for an example of how to do
binding, either way.
 
\end{description}


\subsection{Some statistics collected by dmalloc}
\label{sec:statdmalloc}

A memory consumption problem occured in recent simulations 
(cmcast-[150,200,250].tcl), so we decided to take a closer look at scaling
issue. See page http://www-mash.cs.berkeley.edu/ns/ns-scaling.html
which demostrates the efforts in finding the bottlneck. 
 
The following table summarises the results of investigating the
bottleneck:\\
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|}\hline  
{\em KBytes} &{\em cmcast-50.tcl(217 Links)} &{\em cmcast-100.tcl(950 
Links)}\\\hline 
trace-all &8,084 &28,541\\\hline
turn off trace-all &5,095 &15,465 \\\hline
use array &5,091 &15,459\\\hline 
remove unnecessay variables &5,087 &15,451 \\\hline
on SunOS &5,105 &15,484\\\hline
\end{tabular}
\end{center}
\end{table}


\section{Memory Leaks}
\label{memleak}

This section deals with memory leak problems in \ns, both in Tcl as well
as C/C++.


\subsection{OTcl}
\label{leakTcl}
OTcl, especially TclCL, provides a way to allocate new objects. However,
it does not accordingly provide a garbage collection mechanism for these
allocated objects. This can easily lead to unintentional memory leaks.
Important: tools such as dmalloc and purify are unable to detect this kind
of memory leaks. For example, consider this simple piece of OTcl script: 
\begin{program}
set ns [new Simulator]
for {set i 0} {$i < 500} {incr i} {
        set a [new RandomVariable/Constant]
}
\end{program}
One would expect that the memory usage should stay the same after the
first RandomVariable is allocated. However, because OTcl does not have
garbage collection, when the second RandomVariable is allocated, the
previous one is not freed and hence results in memory leak. Unfortunately,
there is no easy fix for this, because garbage collection of allocated
objects is essentially incompatible with the spirit of Tcl. The only way
to fix this now is to always explicitly free every allocated OTcl object
in your script, in the same way that you take care of malloc-ed object in
C/C++. 


\subsection{C/C++}
\label{leakC}

Another source of memory leak is in C/C++. This is much easier to track
given tools that are specifically designed for this task, e.g., dmalloc
and purify. \ns  has a special target ns-pure to build purified ns
executable. First make sure that the macro PURIFY in the ns Makefile
contains the right -collector for your linker (check purify man page if
you don't know what this is). Then simply type \code{make ns-pure}. See
earlier sections in this chapter on how to use ns with libdmalloc.