Welcome to ZKVM-Perf, a powerful benchmarking tool for ZKVM implementations. This guide covers both automated workflow-based benchmarking and manual benchmarking processes.
- Go to the Actions tab in the ZKVM-Perf repository.
- Click on the "Execute ZKVM-Perf (Matrix)" workflow.
- Click the "Run workflow" button.
- (Optional) Customize parameters or use defaults.
- Click the green "Run workflow" button at the bottom.
- Wait for all matrix jobs to complete.
- Check individual job results and download CSV artifacts for detailed analysis.
The benchmarking process is split into two main workflows:
adhoc-matrix.yml
: Orchestrates the overall benchmarking process.run-on-runner.yml
: Executes the actual benchmarks on EC2 instances.
This workflow sets up the benchmarking environment and triggers individual benchmark runs.
provers
: Provers to use (comma-separated, default: 'sp1')programs
: Programs to benchmark (comma-separated, default: 'loop,fibonacci,tendermint,reth1,reth2')filename
: Filename for the benchmark (default: 'benchmark')trials
: Number of trials to run (default: '1')sp1_ref
: SP1 reference (commit hash or branch name, default: 'dev')additional_params
: Additional parameters as JSON (default: '{"hashfns":"poseidon","shard_sizes":"22"}')
The workflow runs benchmarks on two types of EC2 instances:
- GPU: g6.16xlarge
- CPU: r7i.16xlarge
This workflow is triggered by adhoc-matrix.yml
and runs the actual benchmarks on the specified EC2 instance.
- Sets up the Docker environment.
- Builds the Docker image with the specified SP1 reference.
- Runs the benchmark using the
sweep.py
script. - Uploads the benchmark results as artifacts.
-
Navigate to the Actions Tab Go to the Actions tab in the repository.
-
Select the Workflow Click on "Execute ZKVM-Perf (Matrix)".
-
Configure the Run
- You can use the default settings for a quick start.
- Customize inputs as needed.
-
Start the Benchmark Click "Run workflow".
-
Monitor Progress
- The workflow will start two jobs: one for GPU and one for CPU.
- Each job will trigger a separate
run-on-runner
workflow.
-
Access Results
- Once complete, each job will upload its results as an artifact.
- Download the artifacts to analyze the benchmark data.
- A combined results file will also be available.
-
Install Rust:
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh rustup install nightly
-
Install the SP1 toolchain:
curl -L https://sp1.succinct.xyz | bash source ~/.bashrc sp1up cargo prove --version
-
Install the Risc0 toolchain:
curl -L https://risczero.com/install | bash source ~/.bashrc rzup install cargo risczero --version
-
Install Docker.
-
If using NVIDIA GPUs, install the NVIDIA Container Toolkit.
Note: Run one round of a small program (e.g., Fibonacci) to download the R0 docker image before benchmarking to avoid affecting benchmark times.
Note: On Ubuntu 22.04, you might need to install libssl1.0 for the Risc0 toolchain. Follow these instructions.
To conduct a sweep of the benchmarks:
python3 sweep.py [options]
Available options:
--filename
: Filename for the benchmark (default: "benchmark")--trials
: Number of trials to run (default: 1)--programs
: List of programs to benchmark (choices: loop, fibonacci, tendermint, reth1, reth2)--provers
: List of provers to use (choices: sp1, risc0)--hashfns
: List of hash functions to use (currently only poseidon is supported)--shard-sizes
: List of shard sizes to use--block-1
: Block number for reth1 (default: "17106222")--block-2
: Block number for reth2 (default: "19409768")
To run a single benchmark:
./eval.sh <program> <prover> <hashfn> <shard_size> <filename> [block_number]
Examples:
./eval.sh fibonacci sp1 poseidon 22 benchmark
./eval.sh fibonacci jolt-zkvm poseidon 22 benchmark
./eval.sh fibonacci risc0 poseidon 22 benchmark
./eval.sh reth sp1 poseidon 22 benchmark 19409768
- Each benchmark run produces a CSV file with detailed performance metrics.
- The CSV includes the instance type, allowing for easy comparison between GPU and CPU performance.
- Use the combined results file for a comprehensive view of all benchmarks.
If you encounter issues:
- Check the logs of both the matrix job and the individual runner jobs.
- Ensure your AWS credentials and permissions are correctly set up.
- Verify that the SP1 reference is valid and accessible.
- For GPU jobs, confirm that GPU support is properly configured in the EC2 instance.
For C++ compiler and library issues:
Ubuntu/Debian:
sudo apt update && sudo apt upgrade
sudo apt install build-essential libc6
CentOS/RHEL:
sudo yum update
sudo yum groupinstall "Development Tools"
sudo yum install -y gcc-c++
Fedora:
sudo dnf update
sudo dnf groupinstall "Development Tools"
Setting up NVIDIA:
sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker
We welcome contributions to improve ZKVM-Perf! If you encounter issues or have suggestions:
- Check existing issues in the repository.
- If your issue is new, create a detailed bug report or feature request.
- For code contributions, please submit a pull request with a clear description of your changes.
Happy benchmarking! 🚀