pprof — Visualization and Analysis Tool for Profiling Data

Introduction

pprof is an open-source profiling visualization tool developed by Google. It reads profiling data in the protocol buffer-based pprof format and renders it as interactive flame graphs, directed call graphs, top function tables, and source-level annotations. Originally built for Go's runtime profiler, pprof now supports any profiler that emits the pprof protobuf format, including C++ (gperftools), Java (async-profiler), and Node.js.

What pprof Does

• Reads CPU, memory, goroutine, mutex, and block profiles in the pprof protobuf format
• Renders interactive flame graphs and call graphs in a local web UI
• Generates SVG, PDF, and text reports for sharing and documentation
• Annotates source code and disassembly with per-line profiling data
• Compares two profiles to show regressions or improvements between versions

Architecture Overview

pprof parses a profile protobuf file that contains samples, each sample associating a set of call stack frames with numeric values (CPU time, allocation bytes, etc.). It builds a weighted call graph from these samples and applies user-specified filters (focus, ignore, hide) to narrow the view. The web UI mode starts a local HTTP server that serves an interactive flame graph, a graph visualization (rendered via Graphviz), top tables, and per-line source annotations. The same engine can export static SVG, PDF, or plain text via command-line flags.

Self-Hosting & Configuration

• Install with Go: go install github.com/google/pprof@latest
• Go programs expose profiles at /debug/pprof/ when importing net/http/pprof
• Open interactive web UI: pprof -http=:8080 profile.pb.gz
• Fetch a live profile from a Go server: pprof http://server:port/debug/pprof/profile
• Compare two profiles: pprof -diff_base=old.prof new.prof

Key Features

• Interactive web UI with flame graph, graph view, top table, and source annotation
• Profile diffing to visualize performance changes between two versions
• Supports remote profile fetching from running Go servers over HTTP
• Filtering by function name regex to focus on specific code paths
• Works with any profiler that outputs the pprof protobuf format

Comparison with Similar Tools

• go tool trace — shows goroutine scheduling and system events as a timeline; pprof aggregates CPU and memory samples into call graphs
• Speedscope — web-based flame graph viewer; pprof adds call graph rendering, source annotation, and profile diffing
• FlameGraph (Brendan Gregg) — generates SVG flame graphs from folded stacks; pprof provides an interactive web UI and multiple visualization modes
• py-spy — sampling profiler for Python; pprof is a visualization tool that reads profiles from any runtime that emits pprof format
• Pyroscope / Grafana Pyroscope — continuous profiling platform with storage and querying; pprof is a standalone CLI tool for ad-hoc analysis

FAQ

Q: Is pprof only for Go programs? A: No. Any profiler that outputs the pprof protobuf format works. This includes gperftools (C++), async-profiler (Java), and perf (Linux) with conversion tools.

Q: Do I need Graphviz installed? A: Graphviz is needed for the directed call graph view. Flame graphs and top tables work without it. Install Graphviz with your package manager if you want graph output.

Q: Can I profile a production Go service? A: Yes. Import net/http/pprof in your server, then fetch profiles remotely with pprof http://host:port/debug/pprof/profile?seconds=30.

Q: How do I compare performance between two releases? A: Use pprof -diff_base=baseline.prof current.prof to see a call graph highlighting the delta.

Sources

• https://github.com/google/pprof
• https://pkg.go.dev/net/http/pprof