4

CPU profiling using perf

Alessio Podda edited this page 2024-12-02 07:44:07 +00:00

To profile named using perf, the following steps:

1. Install the prerequisites. E.g. on Fedora:

sudo dnf install perf inferno

inferno is a Rust implementation of the flamegraph utility by Brendan Gregg. Since it is written in a native language, it is faster than the original perl script. If your distribution does not package inferno, you can install the original flamegraph.pl script. E.g.

sudo dnf install perf flamegraph

Replace calls to inferno with flamegraph.pl if necessary.

2. Optional Enable profiling without root access:

echo -1 | sudo tee /proc/sys/kernel/perf_event_paranoid
echo 0  | sudo tee /proc/sys/kernel/nmi_watchdog

3. Ensure you have compiled named with debug symbols. Optionally, set up a prefix

cd ~/Path/to/bind9/checkout
./configure CFLAGS='-O2 -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer -ggdb3' --prefix=/path/to/parent
make clean && make

4. Install named. Normally libtool runs named through a wrapper shell script that can make perf results more difficult to read.

make install

5. Run named under perf.

perf record --compression-level=1 --user-callchains -g --call-graph=dwarf,65528 -e cycles:ppp --output=$OUTPUT_FILE -- $YOUR_NAMED_CMDLINE_GOES_HERE

Below is an explanation of what each of the command line flags do: perf record Runs a command under a sampling profiler. See perf-record(1)

--compression-level=1 Optional Compresses the output file using zstd.

--user-callchains Optional Do not record the kernel.

-g --call-graph=dwarf,65528 Use dwarf symbols and a maximum stack size of 64k. If this option is omitted, inlined functions might not be visualized correctly.

-e cycles:ppp Record stack traces using precise events (:ppp). This option is needed only if you want to look at the assembly.

--output=$OUTPUT_FILE Name of the output file. Otherwise it defaults to perf.data.

6. Generate a flamegraph with inferno:

perf script --input=$INPUT_DATA | inferno-collapse-perf | inferno-flamegraph --width=1920 > $OUTPUT_SVG

You can also instruct inferno to output a "reversed" flamegraph. This is useful to see if there is any leaf function that is used through the codebase and can be optimized:

perf script --input=$INPUT_DATA | inferno-collapse-perf | inferno-flamegraph --reverse --width=1920 > $OUTPUT_SVG