Perf

PERF refers to a suite of performance analysis tools and the underlying perf_events subsystem in the Linux kernel. It provides a unified interface for collecting and analyzing performance data from hardware counters, software counters, and tracing points inside the kernel and in user-space applications.

The perf_events subsystem exposes a set of counters and trace events that can be configured through the perf_event_open system call. Hardware counters measure events such as CPU cycles, instructions, cache references and cache misses, and branch mispredictions, while software counters include context switches and page faults. In addition, perf can collect data from tracepoints, kprobes and uprobes, and can capture call graphs to show where time is spent. The data is typically written to a file (perf.data) and analyzed with user-space tools such as perf report and perf annotate, or viewed in real time with perf record and perf top.

Common workflows include perf stat for a quick summary, perf record to collect detailed samples, and perf report to present a readable summary and, if enabled, a backtrace and source-level mapping. The tool supports multi-core profiling, filtering of events, and generating flame graphs and annotated assembly.

Performs well for performance tuning but has limitations: it requires a Linux kernel with perf_events support and, for certain events, privileged access or a suitable environment (e.g., host or VM permissions). Counter accuracy can vary by CPU model, virtualization, and sampling rate. Overhead from sampling can affect measurements, so results should be interpreted with care and, if possible, validated with multiple runs or alternative methods.