memtrail is a LD_PRELOAD based memory profiler and leak detector for Linux.
There are already many other open-source memory debugging/profiling tools for Linux, several of which are listed on the Links section below, and the most powerful arguably being Valgrind. However, I needed a tool that could quantify leaks and identify memory hogs, for long-running and CPU intensive workloads, which simply run too slow under Valgrind's dynamic binary instrumentation, hence this project.
-
Very little runtime overhead
-
Will immediately output maxmimum memory allocated and total leaked memory at the end of program execution.
-
Text trees show callstacks for memory consuption or leaks
-
Can produce graphs showing flow of memory consumption or leaks
-
Linux
-
Python 3
-
gzip
-
binutils' addr2line
-
For best results (performance/stability) configure libunwind and build as
./configure --disable-cxx-exceptions --disable-debug-frame --disable-block-signals --disable-shared --enable-static --with-picadd set
UNWIND_SRCenvironment variable to where the libunwind source is. -
gprof2dot for graph output
make
Run the application you want to debug as
memtrail record /path/to/application [args...]
and it will generate a record memtrail.data in the current
directory.
View results with
memtrail report --show-maximum
It will produce something like
maximum: 8,960B
-> 45.71% (4,096B, 2x): calloc
| -> 45.71% (4,096B, 2x): test_calloc() [sample.cpp:82]
| -> 45.71% (4,096B, 2x): main [sample.cpp:242]
| -> 45.71% (4,096B, 2x): __libc_start_call_main [sysdeps/nptl/libc_start_call_main.h:58]
| -> 45.71% (4,096B, 2x): call_init [csu/libc-start.c:128]
| -> 45.71% (4,096B, 2x): _start
|
-> 31.43% (2,816B, 4x): malloc
| -> 11.43% (1,024B, 1x): test_malloc() [sample.cpp:57]
| | -> 11.43% (1,024B, 1x): main [sample.cpp:241]
| | -> 11.43% (1,024B, 1x): __libc_start_call_main [sysdeps/nptl/libc_start_call_main.h:58]
| | -> 11.43% (1,024B, 1x): call_init [csu/libc-start.c:128]
| | -> 11.43% (1,024B, 1x): _start
| |
| -> 11.43% (1,024B, 1x): test_realloc() [sample.cpp:103]
| | -> 11.43% (1,024B, 1x): main [sample.cpp:243]
| | -> 11.43% (1,024B, 1x): __libc_start_call_main [sysdeps/nptl/libc_start_call_main.h:58]
| | -> 11.43% (1,024B, 1x): call_init [csu/libc-start.c:128]
| | -> 11.43% (1,024B, 1x): _start
| |
| -> 8.57% (768B, 2x): TestGlobal::TestGlobal() [sample.cpp:212]
| -> 8.57% (768B, 2x): __static_initialization_and_destruction_0(int, int) [sample.cpp:225]
| -> 8.57% (768B, 2x): _GLOBAL__sub_I_leaked [sample.cpp:252]
| -> 8.57% (768B, 2x): call_init [csu/libc-start.c:144]
| -> 8.57% (768B, 2x): _start
|
-> 22.86% (2,048B, 1x): realloc
-> 22.86% (2,048B, 1x): test_realloc() [sample.cpp:106]
-> 22.86% (2,048B, 1x): main [sample.cpp:243]
-> 22.86% (2,048B, 1x): __libc_start_call_main [sysdeps/nptl/libc_start_call_main.h:58]
-> 22.86% (2,048B, 1x): call_init [csu/libc-start.c:128]
-> 22.86% (2,048B, 1x): _start
memtrail.maximum.json written
You can then use gprof2dot.py to obtain graphs highlighting memory leaks or
consumption:
gprof2dot.py -f json memtrail.maximum.json | dot -Tpng -o memtrail.maximum.png
It is also possible to trigger memtrail to take snapshots at specific points by
calling memtrail_snapshot from your code:
#include "memtrail.h"
...
memtrail_snapshot();
Memory debugging:
Memory profiling: