Evaluating Compiler Optimization Impacts on zkVM Performance

This repository contains the code for the paper Evaluating Compiler Optimization Impacts on zkVM Performance, which resulted from my Bachelor's thesis at ETH Zürich.

Getting started

To get started, build the runner:

$ ./scripts/build.sh

Alternatively, with CUDA support, use:

$ ./scripts/build_cuda.sh

After the runner is built, most commands are available via the zkbench cli:

$ ./zkbench.sh --help
Usage: python -m zkbench [OPTIONS] COMMAND [ARGS]...

Options:
  --log-level TEXT
  --log-file TEXT
  --help            Show this message and exit.

Commands:
  asm
  bench
  build
  clean
  plot
  plot-tune
  run
  tune

Rerunning benchmarks

To rerun all benchmarks, first build the binaries:

$ ./zkbench.sh build -j$(nproc)

Next, run the benchmarks via ./zkbench.sh bench. To see all available options, run ./zkbench.sh bench --help.

Autotuning a binary

To explore the optimization space, use the tune subcommand.

Examples:

$ ./zkbench.sh tune --program fibonacci --config ./configs/tune/all-passes.json --metric cycle-count --out test genetic --depth 10 --baseline o3 --mode depth
$ ./zkbench.sh tune --program fibonacci --config ./configs/tune/all-passes-o3-codegen.json --metric cycle-count --out test ffd --resolution 7
$ ./zkbench.sh tune --program fibonacci --config ./configs/tune/all-passes.json --metric cycle-count --out test exhaustive --depth 2

Results

The plots generated from the raw data for the respective experiments in the paper are located in the results directory. All of the raw data is available here. For documentation on how to plot these results, refer to the plotting documentation.

Versions

Environment	Tool	Version
risc0	rust	1.85.0
risc0	cpp	2024.1.5
risc0	r0vm	2.0.0
risc0	cargo-risczero	2.0.0
sp1	sp1-sdk	4.1.4
x86	rust	1.85-x86_64-unknown-linux-gnu