wali-ku/BWLOCK-GPU

Protecting Real-Time GPU Kernels on Integrated CPU-GPU SoC Platforms

Introduction

This repository contains all the scripts needed to reproduce experiments from our ECRTS-18 paper:

Protecting Real-Time GPU Kernels on Integrated CPU-GPU SoC Platforms

Pre-requisites

Hardware

• NVIDIA Jetson TX-2 Board

Software

• CUDA Runtime Library (Version-8.0)
• Linux for Tegra (Version 28.1)
• Python (Version 2.7)
• Git
• Matplotlib

Directory Structure

• kernel: Placeholder directory for hosting Linux for Tegra kernel
  • miscs: Contains required kernel patches
    • diffs
    • devfreq
    • nvgpu
    • tegra-alt
• dynamic_linker: Contains shared library code for intercepting CUDA Runtime API calls
  • lib
• kernel_module: Contains source code of BWLOCK++ kernel module
  • test
    • rt_test
  • src
• benchmarks: Placeholder directory for hosting benchmarks
• scripts: Contains bash and python scripts for automatically running BWLOCK++ experiments
  • graph_scripts
  • individual_figures

Step-by-step Instructions

[Pre-Requisite] We require that the TX-2 be flashed with Jetpack-3.1 before proceeding with the following steps. Please note that flashing the board will erase all data from its internal storage so please back-up any important material you have on your TX-2 board before proceeding any further. All these steps are meant to be executed on the TX-2 board.

1. Launch a bash shell. Install Git.

sudo apt-get install git

2. Clone this repository.

git clone https://github.com/Skidro/BWLOCK-GPU.git

3. Launch a sudo shell.

sudo bash

4. Install BWLOCK++ patched kernel on board (Long Operation - ~1 Hour). All the steps required to do so are automated in this script (RUN-ME.sh) located in BWLOCK-GPU/kernel/ directory. Please note that this requires an active internet connection.

[CAUTION] Please note that running this script will over-write the existing kernel in the /boot/ directory on your TX-2 board!

cd BWLOCK-GPU/kernel
./RUN-ME.sh

5. Reboot the system. Go to the BWLOCK-GPU git repository
6. Setup the benchmarks used in BWLOCK++ evaluation. This is automated by the following script (RUN-ME.sh) in BWLOCK-GPU/benchmarks/ directory. Please note that this requires an active internet connection.

cd benchmarks
./RUN-ME.sh
cd ..

7. [IMPORTANT] Set root path variable (root_path) in this script (functions.sh) located in BWLOCK-GPU/scripts/ directory. The root path is defined as the absolute path in your system where the BWLOCK-GPU repository is located. For example, in our test sytem, the repository is located at /home/nvidia/BWLOCK-GPU which is set as the root path.

8. Launch sudo shell

sudo bash

9. Build all the test materials (benchmarks and kernel module). This is automated by the following script (build_all.sh) located in BWLOCK-GPU/scripts/ directory.

cd scripts
./build_all.sh

10. Run the sanity check experiment (TEST-BWLOCK.sh) located in BWLOCK-GPU/scripts/ directory; to verify that everything is correctly setup in your system. The script takes ~1-minute to complete.

./TEST-BWLOCK.sh

11. Given that your system passes the sanity check, BWLOCK++ is ready to be evaluated on it. However, before proceeding with the final script, install the matplotlib package which is used by the final script for plotting graphs.

apt-get install python-matplotlib

12. Reboot the system. Once the system reboots, go to the directory BWLOCK-GPU/scripts. Relaunch sudo shell and then put the board into maximum performance state using this script (max_perf.sh) located in BWLOCK-GPU/scripts/ directory.

[NOTE] The script (max_perf.sh) will shutdown the GUI on TX-2 board. We suggest that you switch to a TTY terminal (CTRL + ALT + <Function-Key>) before executing this script!

sudo bash
cd BWLOCK-GPU/scripts
./max_perf.sh

13. Run the final evaluation script (RUN-ME.sh) located in BWLOCK-GPU/scripts/ directory; which runs all the experiments from our paper (Protecting Real-Time GPU Kernels on Integrated CPU-GPU SoC Platforms). This will take significant amount of time (Approximately 2 to 3 hours).

./RUN-ME.sh

14. If everything goes alright, the following directory (BWLOCK-GPU/scripts/figures/) should contain all the generated figures.

Figures Reproduced from Paper

By following the instructions mentioned above, the following figures from the paper should get generated in (BWLOCK-GPU/scripts/figures/) directory:

• Figure-1
• Figure-5 (a, b, c)
• Figure-6 (a, b, c)
• Figure-7
• Figure-8
• Figure-9
• Figure-10

Running a Specific Experiment from Paper

In case, one of the individual figures need to be generated from the figures mentioned above, please use the scripts present in this folder (individual_figures) which is located in the BWLOCK-GPU/scripts/ directory. In order to do that, either copy the script for the figure you want to reproduce form the directory (BWLOCK-GPU/scripts/individual_figures/) to the directory (BWLOCK-GPU/scripts/) and then run it OR change the directory to BWLOCK-GPU/scripts and execute the script directly from there (As demonstrated below).

[NOTE] By default, the scripts in individual_figures directory will plot the figure with the existing data in results directory. However, if the data itself needs to be collected, the script should be run with run_the_experiment option.

# Current Directory: BWLOCK-GPU/scripts
# The following command will plot Figure-1 using the existing data in results/motivation folder
./individual_figures/fig1-motivation.sh

# The following command will collect the data for Figure-1 and then plot the graph using the new data
./individual_figures/fig1-motivation.sh run_the_experiment