Example on using the Linux perf profiler
In this note I provide some useful examples on how to use perf. There are lots additional good usage examples at Brendan Gregg’s website.
Identifying sources of allocations
In this example I’ll show you how to identify sources of allocations in
userspace programs, in other words calls to the malloc function. The same idea
can of course be adapted to find calls to any other function.
First add a trace point for malloc:
perf probe -x /lib64/libc.so.6 malloc
Run your program and record allocation events with stack trace:
perf record -e probe_libc:malloc --call-graph dwarf -- sleep 1
View the report:
perf report
In this example with sleep you should get something like this:
# Samples: 30 of event 'probe_libc:malloc'
# Event count (approx.): 30
#
# Children Self Trace output
# ........ ........ ..............
#
100.00% 100.00% (7f1a810c4620)
|
---0x55d6679efbb4
__libc_start_main_alias_2 (inlined)
__libc_start_call_main
|
|--93.33%--0x55d6679ef6b8
| __GI_setlocale (inlined)
| |
| |--50.00%--_nl_find_locale
| | _nl_load_locale_from_archive
| | |
| | |--40.00%--_nl_intern_locale_data
| | | __GI___libc_malloc (inlined)
| | |
| | |--3.33%--_nl_normalize_codeset
| | | __GI___libc_malloc (inlined)
| | |
| | |--3.33%--__GI___strdup (inlined)
| | | __GI___libc_malloc (inlined)
| | |
| | --3.33%--__GI___libc_malloc (inlined)
| |
| |--40.00%--__GI___strdup (inlined)
| | __GI___libc_malloc (inlined)
| |
| --3.33%--new_composite_name
| __GI___libc_malloc (inlined)
|
|--3.33%--0x55d6679ef6cf
| __textdomain
| __GI___strdup (inlined)
| __GI___libc_malloc (inlined)
|
--3.33%--0x55d6679ef6c7
__bindtextdomain
set_binding_values.part.0
__GI___libc_malloc (inlined)