Opera: Achieving Secure and High-performance OLAP via GPU-accelerated Homomorphic Caching

Code Structure

.
├── src                # Source code
├── include            # Header files
├── thirdparty         # Third-party libraries
├── CMakeLists.txt     # CMake configuration file
├── LICENSE            # License file
├── docker-compose.yml # Docker compose file
├── Dockerfile         # Dockerfile
├── cuTFHEpp           # CUDA implementation of TFHEpp library
├── test               # Tests
├── benchmark          # Benchmarks
└── README.md          # README file

Requirements

• C++17/CUDA C++17
• NVIDIA Volta GPU / RTX 30 or later microarchitectures, following are the tested GPUs:
  • NVIDIA A100
• CMake 3.18 or later
• CUDA 11.5 or later
• GPU with 24 GiBs or higher to run the benchmarks

Quick Start

Simply setup the environment with docker compose:

docker compose up -d

Then, build the project in the container:

docker exec -it opera /bin/bash
/# cd app
/app# cmake -B ./build -DCMAKE_CUDA_ARCHITECTURES=80 && make -C build -j

The executable will be generated in the build/bin directory of the project.

Benchmarks

The benchmark can be built with option: -D OPERA_BUILD_BENCHMARKS=ON, and the executable will be located in the build/bin directory.

Available benchmarks:

• tpch_q1: TPC-H Query 1
• tpch_q4: TPC-H Query 4
• tpch_q6: TPC-H Query 6
• tpch_q12: TPC-H Query 12
• tpch_q15: TPC-H Query 15
• tpch_q17: TPC-H Query 17
• tpch_q19: TPC-H Query 19

To check the usage of the benchmark, run the executable with the --help option.

For example, to run the TPC-H Query 6 benchmark with the different options:

// Check the usage
./build/bin/tpch_q6 --help

// Run the benchmark with 16 and 32 rows
./build/bin/tpch_q6 --rows 16 32 --output output.csv

// Run the benchmark with 16 and 32 rows without fast comp
./build/bin/tpch_q6 --rows 16 32 --nofastcomp --output output.csv

// Run the benchmark with 16 and 32 rows without cache
./build/bin/tpch_q6 --rows 16 32 --nocache --output output.csv

// Run the benchmark with 16 and 32 rows without cache and check the result
./build/bin/tpch_q6 --rows 16 32 --nocache --check --output output.csv

Results

The results of the benchmarks are displayed in csv format, and it can be stored in the file specified by the --output option.

The following shows the examples of the result of the TPC-H Query 6 benchmark (in ms):

TPC-H Query 6 without cache:

./build/bin/tpch_q6 --rows 256 1024 4096 16384 --nocache --output opera-NC.csv

TPC-H Query 6 with cache (no fast comp):

./build/bin/tpch_q6 --rows 256 1024 4096 16384 --nofastcomp --output opera-base.csv

TPC-H Query 6 with cache (fast comp):

./build/bin/tpch_q6 --rows 256 1024 4096 16384 --output opera.csv

Cite

If you use Opera for your academic work, you can cite the following paper:

@inproceedings{Opera,
 author      = {Hu, Qi and Chen, Wei and Shen, Tianxiang and Yao, Xin and Zhang, Nicholas and Cui, Heming and Yiu, Siu-Ming},
 booktitle   = {2025 IEEE Symposium on Security and Privacy (SP)}, 
 title       = {Opera: Achieving Secure and High-Performance OLAP with Parallelized Homomorphic Comparisons}, 
 year        = {2025},
 volume      = {},
 number      = {},
 pages       = {2360-2377},
 doi         = {10.1109/SP61157.2025.00173}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Some files contain the modified code from the Phantom version licensed under MIT License.

Some files contain the modified code from the TFHEpp, which is licensed under the Apache License 2.0.