Developer builds contain a performance analysis command called perf that can be used to create an SVG output file similar to bootchart; the chart shows how CPU cycles and I/O wait times are distributed across processes in the system over time.
Below is a short primer in four lessons describing how to generate and view output from perf timechart.
Lesson 1 - a simple example
- Boot Chromium OS, and open a terminal.
- Run this command:
sudo perf timechart record
- Run your workload. A workload isn't necessary if all you want to see is a chart of an idle system. :-)
- When your workload is done, interrupt the process started in step 2 using ^C, or kill -2.
Explanation: Without arguments, perf timechart record runs forever gathering data, until stopped by SIGINT. Note that only SIGINT works; SIGTERM will kill the process without producing the necessary output. When the command completes, you'll see two new files: perf.data and trace.out.
Lesson 2 - how to generate and view the chart
- In the directory where you ran Lesson 1, run this command:
sudo perf timechart
- The output image will be stored in a file named
output.svg. Use scp or some equivalent to copy the file to another system for viewing.
Tips for viewing: Some browsers may have trouble displaying the image. The author of timechart recommends the Inkscape image editor:
Inkscape does a good job of displaying the fine details, but it may be a bit slow for the large timechart images. You should exercise patience when opening, magnifying, or scroling images.
Lesson 3 - how to avoid using SIGINT
- Run this command:
sudo perf timechart record sleep 5
- Run a workload that will finish within 5 seconds; for longer workloads, use a more appropriate sleep time in step 1.
- Generate and view the output as described in Lesson 2.
Explanation: If there are arguments to perf timechart record, the arguments are treated as a command to run as a subprocess of perf. perf gathers data until the process terminates.
If your workload is triggered by a single command, that command can be used in place of 'sleep 5'. Note that if the workload acts as a daemon (that is, forks a child and exits), perf will terminate with the parent terminates; this likely isn't what you'd want.
Lesson 4 - how to get a timechart of system boot
- Add the lines below to
/sbin/chromeos_startup (somewhere after the line that mounts /sys/kernel/debug). This file can be found as src/platform/init/chromeos_startup in your source repository. A patch to apply this change is attached at the bottom of this page.
( cd /tmp ; exec perf timechart record sleep 10 >perf.out 2>&1 ) &
sleep .2
- Add the line below to
/etc/init/boot-complete.conf. This file can be found as src/platform/init/boot-complete.conf in your source repository. A patch to apply this change is attached at the bottom of this page. pkill -2 '^perf$' || true - Boot your system with the changes applied.
- Open a terminal, and run this command:
cd /tmp ; sudo perf timechart
After doing this, you can copy the output.svg file for viewing.
Explanation: This uses the trick from lesson 3, and applies it to early boot time. The sleep .2 command is a hack needed to allow perf to finish initializing before startup scripts that load kernel modules have a chance to run. Without the hack, perf frequently fails to come up. To prevent gathering a full 10 second chart, the boot-complete upstart job kills perf as in lesson 2.
|
ď Richard Barnette, Sep 22, 2010, 2:19 PM
|
