Artefact Evaluation for Parallaft

The repository contains the source code of Parallaft and scripts to reproduce the experiments from the paper "Parallaft: Runtime-based CPU Fault Tolerance via Heterogeneous Parallelism" submitted to CGO 2025.

Repository links

• Latest version of this repository: reproduce-parallaft-paper
• Latest version of Parallaft: parallaft

Prerequisites

• Hardware:
  • CPU:
    • Apple M2 (strongly recommended, works out of the box), or
    • Apple M1 (recommended, but untested), or
    • other Linux-capable Apple Silicon processors (less recommended, needs manual big/little configuration and manual kernel patching, see docs/hardware_support.md), or
    • heterogeneous Intel x86_64 processors (not recommended, lacking energy overhead measurements, and might need manual big/little configuration, see docs/hardware_support.md)
  • RAM: 16GB or more
• Software:
  • OS: Ubuntu 24.04 (on x86_64) or Ubuntu Asahi 24.04 (on Apple Silicon aarch64)
  • Benchmark: SPEC CPU 2006 (obtained separately)
  • Runtime environment: Available RAM and swap combined should be at least 24GB

Installation

To clone the repository and install dependencies, run:

$ git clone https://github.com/CompArchCam/reproduce-parallaft-paper --recursive
$ ./scripts/deps.sh

Log out and log back in to enable Docker access without sudo.

To build Parallaft binary, run:

$ ./scripts/build_app.sh

To install SPEC CPU 2006, first obtain a copy of SPEC CPU 2006 ISO and place cpu2006-1.2.iso at the root of the artefact package. Then execute:

$ ./scripts/install_spec06.sh

Additional steps for Apple Silicon aarch64

On Apple Silicon aarch64, we need a patched Linux kernel to access power sensor readings (for measuring energy consumption in our experiments) and for Parallaft to monitor dirty pages. To do that, run

$ ./scripts/build_kernel.sh
$ sudo reboot
$ ./scripts/check_kernel.sh

If everything is okay, you should see the following.

Checking kernel version...
Kernel version is OK
Checking power sensor...
Power sensor is OK

Our patch only supports reading power sensors from Apple M1 and M2 processors (non-Pro/Max/Ultra models). If you are not using one of these, check docs/hardware_support.md for guide to add support for your processor.

Workflow

Running all experiments

In our experiments, the following runs are performed on all SPEC CPU 2006 int and fp benchmarks.

• A baseline run, to get execution time and CPU time without Parallaft or RAFT.
• (aarch64 only) A baseline energy-consumption profiling run, to get baseline energy consumption without Parallaft or RAFT.
• A Parallaft run.
• A RAFT run.

To run all experiments, run:

$ ./scripts/run.sh

Raw results (*.stats.txt) will be available under spec06/releval/run/*/result.

Plotting results

Our experiments reproduce the following plots.

• Performance overhead of Parallaft and RAFT (Figure 5).
• Performance-overhead breakdown of Parallaft (Figure 6).
• Energy overhead of Parallaft and RAFT (Figure 7). This result will only be available on Apple Silicon aarch64 platforms.

To plot them, run:

$ ./scripts/plot.sh

Plots will be available under plots directory. On an Apple M2, they should look broadly similar to the figures in our paper.

Customization

• Running a subset of benchmarks or experiments: Modify BENCHMARKS and EXPERIMENTS in scripts/run.sh.
• Tuning parameters: Adjust PARALLAFT_CHECKPOINT_PERIOD in scripts/run.sh.

Running an arbitrary program under Parallaft

$ ./bin/parallaft --config ./app/parallaft/configs/apple_m2_fixed_interval.yml -- path/to/your/program arg1 arg2

When the execution finishes, Parallaft dumps some statistics. The key ones are:

• timing.all_wall_time: Wall time elapsed to finish the program execution, including the waiting time for outstanding checkers to finish after the main finishes.
• timing.main_wall_time: Wall time elapsed to finish the main program execution, excluding the waiting time for checkers.
• timing.main_{user,sys}_time: User/system time used by the main program execution.
• hwmon.macsmc_hwmon/*: (Apple Silicon only) Energy used for different components of the SoC during the program execution.
• counter.checkpoint_count: Number of checkpoints taken, including checkpoints taken to handle certain mmap syscalls and to slice the program execution for checker parallelism.
• fixed_interval_slicer.nr_slices: Number of checkpoints taken to slice the program execution after reaching a specified checkpoint period.

Use another config to run on a Intel processor, e.g. for Intel Core i7-12700, use intel_12700_fixed_interval.yml.

Uninstalling

Uninstalling the custom Apple Silicon aarch64 kernel

$ sudo -i
$ rm /boot/*-6.10.5-asahi-parallaft
$ rm -r /lib/firmware/6.10.5-asahi-parallaft
$ update-grub
$ update-m1n1