perf

perf [stat|record|report|top|list] [options] [command]

perf is the official Linux kernel profiling tool. It provides access to hardware performance counters, tracepoints, and software performance events. It's essential for identifying performance bottlenecks in applications and the kernel.The tool can measure CPU cycles, cache misses, branch mispredictions, and many other low-level events. It supports both sampling-based profiling and event counting.

-p, --pid _pid_

Profile existing process by PID

Select specific events to measure

Enable call-graph (stack chain/backtrace) recording

System-wide collection from all CPUs

Profile at specified frequency (samples per second)

Output file name

Collect a sample after every count events

stat

Run command and gather performance counter statistics

Run command and record profile into perf.data

Read perf.data and display profile

System-wide real-time profiling

List available performance events

Read perf.data and display annotated code

Read perf.data and output trace in scripted format

Run built-in microbenchmarks for scheduler, memory, etc.

Trace system calls (similar to strace)

Analyze scheduler latencies and behavior

Compare two perf.data files

Requires root or kernel.perfeventparanoid sysctl adjustment for many operations. Hardware counter availability varies by CPU. Profile data can be large. Some features require specific kernel configuration.

perf was developed as part of the Linux kernel, primarily by Ingo Molnar and Thomas Gleixner. It replaced the older OProfile tool and became the standard Linux profiling interface, integrated directly into the kernel source tree.

strace(1), ltrace(1), valgrind(1), gprof(1)