CUDA integration with PostgreSQL

This project provides examples of how to integrate CUDA with PostgreSQL. It includes the following: a vector addition extension computed on the GPU, a Set Returning Function (SRF) that computes the maximum number of each time interval using a parallel reduction algorithm on the GPU, and an implementation of a CUDA kernel that computes the heart rate of a resident based on their bed sensor data. The heart rate estimation kernel results are the ones presented in the submitted paper. The project also includes a Python script that can be used to reproduce the results of the paper by inserting dummy data into the database and running the heart rate estimation function over different time intervals. Lastly, the project includes a script that can be used to generate a plot of the results of the heart rate estimation function over different time intervals.

Prerequisites

• Python 3.6 or higher installation instructions
• Valid installation of PostgreSQL (version 10 or higher)
• Having a superuser role in PostgreSQL (i.e. postgres user in Linux has superuser role by default).
• Development libraries and headers for C language backend for PostgreSQL
  • sudo apt install postgresql-server-dev-XX where XX is the version of PostgreSQL
• CUDA Toolkit (version 10.1 or higher)

Repository Installation

git clone https://github.com/MU-HPDI/postgres-extensions-with-cuda.git
cd postgres-extensions-with-cuda

Usage

1. Make sure that the PostgreSQL server is running.
2. Check valid GPU devices: nvidia-smi Example output:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 470.161.03   Driver Version: 470.161.03   CUDA Version: 11.4     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  Off  | 00000000:0A:00.0 Off |                  N/A |
|  0%   37C    P8     2W / 225W |    159MiB /  7981MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1229      G   /usr/lib/xorg/Xorg                 36MiB |
|    0   N/A  N/A      1475      G   /usr/bin/gnome-shell                9MiB |
|    0   N/A  N/A    159183      C   ...ied 127.0.0.1(60178) idle      109MiB |
+-----------------------------------------------------------------------------+

3. Check the include PostgreSQL directory: pg_config --includedir-server
4. Change the PG_SERVER_DIR variable in the Makefile file to the directory returned if wanting to use a different version of PostgreSQL. (pg_config --includedir-server returns the directory for the latest version of the installed PostgreSQL.)
5. Create a build directory: mkdir build
6. Set the required environment variables in a .env file from the .env.example file.
7. Insert dummy data into the database: make insert
8. Run the project: make clean extension

PostgreSQL Functions

Once the project is running, you can execute the following functions:

This function returns the sum of two integer vectors.

SELECT vector_addition_cuda(ARRAY[1, 2, 3, 4, 5, 6, 7, 8, 9, 10], ARRAY[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]);

This function returns the maximum number of each time interval using a parallel reduction algorithm on the GPU.

SELECT * FROM max_reduction_cuda('2023-04-08 22:00:00'::TIMESTAMP, '2023-04-09 22:00:00'::TIMESTAMP);

This function returns the heart rate of a resident based on their bed sensor data.

SELECT * FROM heart_rate_estimation('bed_data', '2021-05-29 00:00:00'::TIMESTAMP, '2021-06-07 00:00:00'::TIMESTAMP, 100, 'GPU', '1.0');

Environment Variables

The following environment variables are required to run the project:

• PGUSER: The PostgreSQL username
• PGPASSWORD: The PostgreSQL password
• PGHOST: The PostgreSQL server host name
• PGPORT: The PostgreSQL server port number
• PGDATABASE: The PostgreSQL database name

Reproducibility

The following steps can be used to reproduce the results of this project (i.e. Fig. 5 from the paper). In this step, we are going to get the heart rate estimation over different time intervals using the GPU and CPU (1 to 1800 minutes of bed sensor data). The experiemnt can take a minute to run and generate the plot. The following steps assume that the project has been installed and all data has been inserted into the database. (See the Repository Installation and Usage sections for more details.):

1. Run the following command:

make plot

2. The plot will be saved in the scripts directory.

Note: If the plot is already generated, you need to change the experiment_version variable in the scripts/plots.py file to perform a new experiment.

License

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

Resources

• CUDA Toolkit Documentation
• Thrust Documentation
• PostgreSQL Documentation
• SPI Documentation
• PostgreSQL C API Documentation