Perf performance debugging

perf yes linux One of the system debugging tools , It can be sampled at a certain frequency ( Default 4000Hz) Monitor the operation of functions in the specified process , So as to analyze the performance bottleneck of the program , And then targeted optimization .

Tool installation

perf Not at all linux The system comes with it by default , Need to install linux-tools Toolset , such as ：

sudo apt install linux-tools-common linux-tools-generic linux-tools-`uname -r`

Or compile it from the source code ：

sudo apt install build-essential git flex bison
git clone --depth 1 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
cd linux/tools/perf
make
sudo cp perf /usr/bin

If your system kernel has been upgraded by yourself , Then you can only compile from the source code , because apt The source only has the default kernel of the current system linux-tools-xxx

perf Usage method

First, you need to run the tested program , And get its PID, Suppose the program name is my_awesome_app, Then get PID The order is ：

MY_PID=$(ps -ef | grep my_awesome_app | grep -v 'grep' | awk '{print $2}')

Then start perf monitoring ：

perf record -F 4000 -g -p $MY_PID -- sleep 80

among -F <freq> Specify the sampling frequency ,-g It means recording function call stack ,-p <PID> Specify monitored PID,-- sleep <secs> Specify the monitoring duration , Please check perf record --help

perf After monitoring, it will be generated in the current path perf.data file , perform perf report -n --stdio Will read perf.data And display the results on the terminal .

Flame chart

perf report Analyzing performance is still inconvenient , It can be converted into a flame diagram to view , Tools to address ：FlameGraph, The specific usage has been clearly written in the warehouse .

Reference material

• How to read flame chart ？
• Linux perf Examples