Michael Paquier - PostgreSQL committer

The Linux kernel comes with a performance analysis tool called perf that can even be used on production servers to analyze the call stack of a server, or a process. This is useful to find the bottlenecks of a query processing or simply of a server running and improve those.

It is usually found in the package “perf”, at least it is the case of RHEL, CentOS and Linux.

/proc/sys/kernel/perf_event_paranoid can be used to restrict access to the performance counters.

• 2, allow only user-space measurements
• 1, allow both kernel and user measurements (default)
• 0, allow access to CPU-specific data
• -1, no paranoid at all

Record information

Profile system as long as needed, can be cancelled with Ctl-C:

perf record -a -g

Profile system for a given period of time:

perf record -a -g -s sleep $TIME_IN_SECONDS

For the duration of a command:

perf record -a -g -s -- $COMMAND

Different things can be recorded:

• -a to profile the whole system
• -p $PID to profile only the given PID
• -u $USER to profile only the given user

Reports are saved by default in $HOME/perf.data. Old reports are renamed as $HOME/perf.data.old.

Real-time measurement

This is done by perf top, like system-wide profiling:

perf top

Profiling without accumulating stats:

perf top -z

View records

View profile recorded:

perf report -n

With a graph:

perf report -g