Orthrus: Efficient and Timely Detection of Silent User Data Corruption in the Cloud with Resource-Adaptive Computation Validation

Orthrus is a system for the timely detection of silent user-data corruption (SDC) caused by post-installation CPU errors. Orthrus enables high-coverage protection of user-data in the cloud with minimal performance impact.

For more details, please refer to our SOSP 2025 paper:

Orthrus: Efficient and Timely Detection of Silent User Data Corruption in the Cloud with Resource-Adaptive Computation Validation

---

Table of Contents

• Overview
• Environment and Dependencies
• Getting Started
• Evaluation
  • Performance Testing
  • Error Injection Tests
• Related Repositories
• Citation

---

Overview

This repository contains the research artifact for evaluating the performance of the Orthrus Runtime.

For the automated testing framework of Fault-Injection (SDC errors Injection using LLVM compilers), please refer to the FaultInjection repository. We are continuing to develop that repository into an easy-to-use automated test platform for fault injection.

Environment and Dependencies

Orthrus is evaluated on Ubuntu 18.04 and Ubuntu 20.04 (recommended).

• CPU: Minimum 48 cores required for testing scripts.
• CMake: Version 3.20 or higher.
• Detailed requirements can be found in docs/prerequisite.md.

Getting Started

1. Environment Setup

On Ubuntu 20.04, you can initialize the environment using the provided script:

sudo ./init.sh
source env.sh

Alternatively, you can use Docker Compose to run evaluations without manual setup.

2. Dataset Preparation

To calculate the detection rate, you need pre-generated fault injection results.

1. Download the file fault_injection.tar.gz from: OneDrive Link
2. Extract it to the datasets/ directory.

3. Running Evaluations

You can run the full evaluation suite using one of the following methods:

Method A: Native Environment

just test-all

Method B: Docker Compose

docker-compose run test-all

Performance results will be saved in the results/ folder.

---

Evaluation Details

Performance Testing

This section corresponds to Figures 6-9 in the paper. The tests cover: Memcached, Masstree, Phoenix, and LSMTree.

• The complete performance evaluation takes approximately 7 hours.
• For details of individual tests, please refer to:
  • Memcached Evaluation
  • Masstree Evaluation
  • Phoenix Evaluation
  • LSMTree Evaluation

Error Injection Tests

This section corresponds to Table 2 in the paper.

• Full Test: A complete error injection test may take over 30 hours. Results in the paper are based on the full test.
• Fast Check: We provide a partial test for quick verification (~2-3 hours).
• More details can be found in Table 2 Fast Check and Table 2 Full.

Related Repositories

• FaultInjection: Automated testing framework for SDC error injection using LLVM compilers.

Citation

If you use Orthrus in your research, please cite our SOSP 2025 paper:

@inproceedings{liu2025orthrus,
  title={Orthrus: Efficient and Timely Detection of Silent User Data Corruption in the Cloud with Resource-Adaptive Computation Validation},
  author={Liu, Chenxiao and Zhu, Zhenting and Li, Quanxi and Xia, Yanwen and Qiao, Yifan and Deng, Xiangyun and Lu, Youyou and Xie, Tao and Cui, Huimin and Du, Zidong and others},
  booktitle={Proceedings of the ACM SIGOPS 31st Symposium on Operating Systems Principles},
  pages={286--304},
  year={2025}
}