Real-Time Scheduling with NVIDIA Jetson TX2

NVIDIA Jetson TX2 Configuration

• Visit Nvidia L4T download directory for the list of packages
  Latest stable release at the time for Jetson TX2 platform is "L4T 28.1 - Production Version"
  Download all packages to $HOME/nvidia, rest of the document assumes that

1. NVIDIA_Tegra_Linux_Driver_Package.tar : Documentation
2. Tegra186_Linux_R28.1.0_aarch64.tbz2 : Jetson TX2 64-bit Driver Package
3. Tegra_Linux_Sample-Root-Filesystem_R28.1.0_aarch64.tbz2 : Sample Root File System
4. gcc-4.8.5-aarch64.tgz : GCC 4.8.5 Tool Chain for 64-bit BSP (which contains gcc-4.8.5 Jetson TX2 release for cross-compilation)
5. source_release.tbz2 Source Packages : (which contains Linux kernel sources for Jetson TX2 platform)

Optionally you may download JetPack to extract CUDA .deb packages or download them directly
6. JetPack-L4T-3.1-linux-x64.run: All-in-one package Jetson SDK containing CUDA, TensorRT, cuDNN, VisionWorks/OpenCV4Tegra, Samples/Documentation

• How to build an image and flash into eMMC
• How to configure a fresh install
• How to develop and update board software

Real-Time Patches and L4T Kernel

• Real-Time Linux Project contains necessary information to add real-time capabilities to vanilla kernels
• How to patch and build a real-time Jetson kernel

Performance Comparison (RT & non-RT)

Below are a selection of popular benchmark suites selected to probe different aspects of CPU/GPU hybrid systems

• Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by Berkeley AI Research (BAIR)/The Berkeley Vision and Learning Center (BVLC) and community contributors
• Periodic Task Releaser is a collection of CUDA programs intended to measure interference between GPU processes
• CUDA Memory Experiments are simple programs that run memory experiments on CUDA
• Mixbench is a GPU benchmark tool for evaluating GPUs on mixed operational intensity kernels (CUDA, OpenCL, HIP)
• Rodinia Benchmark Suite is a collection of parallel programs which targets heterogeneous computing platforms with both multicore CPUs and GPUs

# to build all benchmarks
$ make -C benchmarks all

# or to build a specific one
$ make -C benchmarks caffe_all
$ make -C benchmarks periodictaskreleaser_all
$ make -C benchmarks cudamemoryexperiments_all
$ make -C benchmarks mixbench_all
$ make -C benchmarks rodinia_all


# to clean all benchmarks
$ make -C benchmarks clean

For details on how to build and run benchmark suites
For details on system caps and throttle script