plasma-umass/coz

Coz: Finding Code that Counts with Causal Profiling

by Charlie Curtsinger and Emery Berger

Coz is a profiler for native code (C/C++/Rust) that unlocks optimization opportunities missed by traditional profilers. Coz employs a novel technique called causal profiling that measures optimization potential. It predicts what the impact of optimizing code will have on overall throughput or latency.

Profiles generated by Coz show the "bang for buck" of optimizing a line of code in an application. In the below profile, almost every effort to optimize the performance of this line of code directly leads to an increase in overall performance, making it an excellent candidate for optimization efforts.

Coz's measurement matches developers' assumptions about profilers: that optimizing highly-ranked code will have the greatest impact on performance. Causal profiling measures optimization potential for serial, parallel, and asynchronous programs without instrumentation of special handling for library calls and concurrency primitives. Instead, a causal profiler uses performance experiments to predict the effect of optimizations. This allows the profiler to establish causality: "optimizing function X will have effect Y," exactly the measurement developers had assumed they were getting all along.

Full details of Coz are available in our paper, Coz: Finding Code that Counts with Causal Profiling (pdf), SOSP 2015, October 2015 (recipient of a Best Paper Award).

Installation

Pre-built Packages (Recommended)

Download the latest release for your platform from the GitHub Releases page.

Debian/Ubuntu (.deb):

# Download the .deb for your architecture (amd64 or arm64)
sudo dpkg -i coz_VERSION_amd64.deb

Fedora/RHEL/CentOS (.rpm):

# Download the .rpm for your architecture (x86_64 or aarch64)
sudo rpm -i coz-VERSION-1.x86_64.rpm

Generic Linux (tarball):

tar xzf coz-VERSION-linux-x86_64.tar.gz
cd coz-VERSION-linux-x86_64
sudo ./install.sh        # Installs to /usr/local by default
sudo ldconfig

Requirements

Coz works on Linux systems (running version 2.6.32 or later, with support for the perf_event_open system call) and macOS (using Apple's kperf framework). Both platforms require a Python 3.x interpreter.

macOS Note: The macOS port uses Apple's private kperf framework for sampling. This requires either running with elevated privileges or adjusting System Integrity Protection settings. The kperf API is undocumented and may change in future macOS versions.

Libraries/Wrappers

By default, Coz works for C, C++, and Rust programs. It has been ported or has wrappers for several other languages, listed below:

Building Coz From Source

Install build prerequisites

On Debian/Ubuntu:

sudo apt-get install -y build-essential cmake pkg-config

libelfin is fetched automatically during the build, so no additional dependencies are required.

Configure and build

Use the standard out-of-source workflow (shown with build/, but any directory works):

cmake -S . -B build          # Configure (defaults to Release with debug info)
cmake --build build -j       # Build libcoz and the CLI
sudo cmake --install build   # Install to /usr/local (or CMAKE_INSTALL_PREFIX)
sudo ldconfig                # Update shared library cache

Before running Coz on Linux, relax perf_event_paranoid so sampling works:

sudo sh -c 'echo 1 >/proc/sys/kernel/perf_event_paranoid'

Building the Benchmarks

The benchmark suite is off by default because it pulls in extra dependencies. Enable it when configuring:

cmake -S . -B build-bench -DBUILD_BENCHMARKS=ON
cmake --build build-bench -j

When BUILD_BENCHMARKS is set, CMake automatically switches the build type to RelWithDebInfo (or keeps Debug) so DWARF line tables are available. Benchmark binaries live under build-bench/benchmarks/<name>/.

A number of the benchmarks are from the Phoenix benchmark suite, and several require data files. These are available for download via links in the README from the Phoenix repository.

Viewer

After profiling, run coz plot to automatically open your results in the browser. To view a specific profile, use coz plot -i /path/to/profile.coz.

You may need to adjust the "Minimum Points" slider to see results if the profile has limited data.

AI-Powered Optimization Suggestions

The viewer includes an AI assistant that analyzes causal profiling results and suggests concrete optimizations. Click the magic wand icon on any plot to get context-aware suggestions based on the profiling data and source code.

Supported LLM providers:

• Anthropic (Claude) — set ANTHROPIC_API_KEY
• OpenAI (GPT-4o, o3, etc.) — set OPENAI_API_KEY
• Amazon Bedrock — set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY
• Ollama — local models, no API key required

The viewer dynamically fetches available models from each provider and caches them locally. Provider settings, API keys, and model selections persist across sessions via cookies.

Using Coz

Using Coz requires a small amount of setup, but you can jump ahead to the section on the included sample applications in this repository if you want to try Coz right away.

To run your program with Coz, you will need to build it with debug information (-g). Coz now supports modern DWARF versions (including DWARF 5), so you can use your compiler's default debug format. You do not need to include debug symbols in the main executable: coz uses the same procedure as gdb to locate debug information for stripped binaries.

Once you have your program built with debug information, you can run it with Coz using the command coz run {coz options} --- {program name and arguments}. But, to produce a useful profile you need to decide which part(s) of the application you want to speed up by specifying one or more progress points.

Profiling Modes

Coz departs from conventional profiling by making it possible to view the effect of optimizations on both throughput and latency. To profile throughput, you must specify a progress point. To profile latency, you must specify a pair of progress points.

Throughput Profiling: Specifying Progress Points

To profile throughput you must indicate a line in the code that corresponds to the end of a unit of work. For example, a progress point could be the point at which a transaction concludes, when a web page finishes rendering, or when a query completes. Coz then measures the rate of visits to each progress point to determine any potential optimization's effect on throughput.

To place a progress point, include coz.h (under the include directory in this repository) and add the COZ_PROGRESS macro to at least one line you would like to execute more frequently. Don't forget to link your program with libdl: use the -ldl option.

By default, Coz uses the source file and line number as the name for your progress points. If you use COZ_PROGRESS_NAMED("name for progress point") instead, you can provide an informative name for your progress points. This also allows you to mark multiple source locations that correspond to the same progress point.

Latency Profiling: Specifying Progress Points

To profile latency, you must place two progress points that correspond to the start and end of an event of interest, such as when a transaction begins and completes. Simply mark the beginning of a transaction with the COZ_BEGIN("transaction name") macro, and the end with the COZ_END("transaction name") macro. Unlike regular progress points, you always need to specify a name for your latency progress points. Don't forget to link your program with libdl: use the -ldl option.

When coz tests a hypothetical optimization it will report the effect of that optimization on the average latency between these two points. Coz can track this information without any knowledge of individual transactions thanks to Little's Law.

AI-Suggested Progress Points (coz suggest-points)

If you're new to Coz or working in an unfamiliar codebase, the hardest part is deciding where to place progress points. The coz suggest-points subcommand uses an LLM agent to read your source, identify what counts as a unit of work, and propose concrete COZ_PROGRESS_NAMED / COZ_BEGIN / COZ_END placements. Each proposal is shown with a rationale and a unified diff; nothing is written until you confirm.

export ANTHROPIC_API_KEY=...

coz suggest-points src/                         # explore src/, then prompt to apply
coz suggest-points --dry-run src/               # print diffs only, no prompt
coz suggest-points --apply src/                 # skip the prompt, apply everything
coz suggest-points --kind latency src/          # only propose COZ_BEGIN/COZ_END pairs
coz suggest-points --hint "HTTP server" src/    # give the agent a domain hint

The agent has read-only tools (list_files, read_file, grep) scoped to the paths you pass, and emits one proposal per call. When you accept a proposal, the macro is inserted at the chosen line (matching surrounding indentation), #include "coz.h" is added if missing, and the original file is backed up to *.coz.bak. Latency points are only applied in matched BEGIN/END pairs. Files that already contain Coz macros are left alone.

Choosing an LLM provider

coz suggest-points supports the same providers as the viewer's optimization assistant. Pick one with --provider; the agentic tool-use loop is used on all of them, so results are comparable across providers (quality depends on how well the chosen model follows tool-use instructions).

Examples:

# Anthropic (default)
export ANTHROPIC_API_KEY=...
coz suggest-points src/

# OpenAI
export OPENAI_API_KEY=...
coz suggest-points --provider openai --model gpt-4o src/

# Amazon Bedrock — uses the boto3 credential chain
export AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... AWS_REGION=us-west-2
coz suggest-points --provider bedrock \
                   --model anthropic.claude-opus-4-20250514-v1:0 src/

# Ollama on a local server (no API key)
coz suggest-points --provider ollama --model llama3.1 src/

You can also pass --api-key <key> inline for Anthropic or OpenAI instead of exporting an env var, and --model <id> to select a specific model on any provider. See coz suggest-points --help for the full option list (--include, --exclude, --max-points, --region, --ollama-host, etc.).

Specifying Progress Points on the Command Line

Coz has command line options to specify progress points when profiling the application instead of modifying its source. This feature is currently disabled because it did not work particularly well. Adding support for better command line-specified progress points is planned in the near future.

Processing Results

Run coz plot to view your profile in the browser. Use coz plot --text for terminal output, or coz plot --text --verbose for detailed data points.

Sample Applications

The benchmarks/ directory includes several small programs with progress points already wired up. Once you configure with -DBUILD_BENCHMARKS=ON (see above), you can run them straight from the build tree:

./build-bench/benchmarks/toy/toy
coz run --- ./build-bench/benchmarks/toy/toy

These programs may need several runs before Coz accumulates enough samples to emit a useful profile. Run coz plot to view the results.

CMake

When you install coz it installs a cmake config file. To add coz to a cmake project simply use the command find_package(coz-profiler). This will import a target for the library and includes called coz::coz and a target for the coz binary coz::profiler. For guidance on how to use these targets refer to the CMake documentation.

Limitations

Coz currently does not support interpreted or JIT-compiled languages such as Python, Ruby, or JavaScript. Interpreted languages will likely not be supported at any point, but support for JIT-compiled languages that produce debug information could be added in the future.

License

All source code is licensed under the BSD 2-clause license unless otherwise indicated. See LICENSE.md for details.

Sample applications (in the benchmarks directory) include several Phoenix programs and pbzip2, which are licensed separately and included with this release for convenience.