Home

Awesome

Memory Analysis Tool

Build Status Download License Commits

MAT is a simple memory analysis tool intended to help understand where the memory is used in a program.

The tool works by using a small shared library that can be loaded by using the LD_PRELOAD dynamic linker option. The shared library collects memory allocation events and generates an event file that can be analyzed by MAT tool.

From the event file, the tool is able to provide useful information. This includes the details about the memory allocation (size, address), the complete stack frame where the memory allocation was made, the timestamp and thread information.

MAT was created to answer these simple questions:

It is similar to the Massif heap profiler provided by a Valgrind plugin.

A first version of MAT existed back in 1994 but it was written in C++ when Valgrind was not available.

Version 1.2 - Under development

Version 1.1 - Apr 2021

Using git

The project uses git submodules to integrate several other projects. To get all the sources, use the following commands:

git clone --recursive https://github.com/stcarrez/mat.git
cd mat

Building mat

The package is composed of three separate components:

libmat and matl are written in C and mat is written in Ada which requires the GNAT Ada compiler. By default the mat component is not built. If you only need libmat and matl, configure and build as follows:

./configure
make

To build a 32-bit or 64-bit version of the shared library you may use:

  CC="/usr/bin/gcc -m32" ./configure --disable-mat
  make

or

  CC="/usr/bin/gcc -m64" ./configure --disable-mat
  make

If you're using a cross compilation environment, you should indicate to the configure your target host. For example to build for a remote mips system, use:

  ./configure --host="mips-uclibc-linux" --target=mips-uclibc-linux --disable-mat
  make

To build the mat analysis tool, you must have installed the following components on your system:

On Debian-based systems, you may have to install the following packages:

sudo apt-get install gnat gprbuild binutils-dev libiberty-dev libreadline-dev

If you have not installed Ada Utility Library and Ada Bfd Library on your system, you can configure and build by using:

  ./configure --enable-ada-util --enable-ada-bfd
  make

Instrumenting your application

You can instrument your application passively by recording all events and looking at the memory allocation after the program has stopped. It is also possible to instrument dynamically while the application is running. Both methods have they advantages.

Passive instrumentation

You can instrument the memory allocation by using the matl launcher.

  matl -o name my-program

While the program runs and the libmat.so collect events, it generates a file name-<pid>.mat.

Start mat with the generated file:

  mat name-xxx.mat

Once the memory events are loaded, you can use the interactive commands to look at the events. The first commands you may use are info, timeline and sizes as they give a short summary and analysis of the events.

Dynamic instrumentation

The dynamic instrumentation requires that the mat analyser is started in the server mode: it is started first, before the application to analyse. The server is activated by the -s option. It listens to the TCP/IP port 4606 and then enter in the interactive mode:

mat -s

Then, you can launch your application through the same matl launcher but you will specify the host name to connect:

  matl -s localhost my-program

Embedded systems

On embedded systems, you only need to build the libmat.so and matl parts. For Mips and ARM, for the backtrace to work, you should compile your program with the -funwind-tables gcc option. Instrument your program and copy the generated .mat files on your Linux host. Make sure your program is not stripped and available to the mat program to get the symbols (use the -s path option if necessary).

Documentation

Articles