Red Hat Training

A Red Hat training course is available for RHEL 8

Chapter 20. Investigating busy CPUs with perf

When investigating performance issues on a system, you can use the perf tool to identify and monitor the busiest CPUs in order to focus your efforts.

20.1. Displaying which CPU events were counted on with perf stat

You can use perf stat to display which CPU events were counted on by disabling CPU count aggregation. You must count events in system-wide mode by using the -a flag in order to use this functionality.

Prerequisites

  • You have the perf user space tool installed as described in Installing perf.

Procedure

  • Count the events with CPU count aggregation disabled:

    # perf stat -a -A sleep seconds

    The previous example displays counts of a default set of common hardware and software events recorded over a time period of seconds seconds, as dictated by using the sleep command, over each individual CPU in ascending order, starting with CPU0. As such, it may be useful to specify an event such as cycles:

    # perf stat -a -A -e cycles sleep seconds

20.2. Displaying which CPU samples were taken on with perf report

The perf record command samples performance data and stores this data in a perf.data file which can be read with the perf report command. The perf record command always records which CPU samples were taken on. You can configure perf report to display this information.

Prerequisites

  • You have the perf user space tool installed as described in Installing perf.
  • There is a perf.data file created with perf record in the current directory. If the perf.data file was created with root access, you need to run perf report with root access too.

Procedure

  • Display the contents of the perf.data file for further analysis while sorting by CPU:

    # perf report --sort cpu
    • You can sort by CPU and command to display more detailed information about where CPU time is being spent:

      # perf report --sort cpu,comm

      This example will list commands from all monitored CPUs by total overhead in descending order of overhead usage and identify the CPU the command was executed on.

20.3. Displaying specific CPUs during profiling with perf top

You can configure perf top to display specific CPUs and their relative usage while profiling your system in real time.

Prerequisites

  • You have the perf user space tool installed as described in Installing perf.

Procedure

  • Start the perf top interface while sorting by CPU:

    # perf top --sort cpu

    This example will list CPUs and their respective overhead in descending order of overhead usage in real time.

    • You can sort by CPU and command for more detailed information of where CPU time is being spent:

      # perf top --sort cpu,comm

      This example will list commands by total overhead in descending order of overhead usage and identify the CPU the command was executed on in real time.

20.4. Monitoring specific CPUs with perf record and perf report

You can configure perf record to only sample specific CPUs of interest and analyze the generated perf.data file with perf report for further analysis.

Prerequisites

  • You have the perf user space tool installed as described in Installing perf.

Procedure

  1. Sample and record the performance data in the specific CPU’s, generating a perf.data file:

    • Using a comma separated list of CPUs:

      # perf record -C 0,1 sleep seconds

      The previous example samples and records data in CPUs 0 and 1 for a period of seconds seconds as dictated by the use of the sleep command.

    • Using a range of CPUs:

      # perf record -C 0-2 sleep seconds

      The previous example samples and records data in all CPUs from CPU 0 to 2 for a period of seconds seconds as dictated by the use of the sleep command.

  2. Display the contents of the perf.data file for further analysis:

    # perf report

    This example will display the contents of perf.data. If you are monitoring several CPUs and want to know which CPU data was sampled on, see Displaying which CPU samples were taken on with perf report.