Skip to content

Latest commit

 

History

History
271 lines (213 loc) · 18.7 KB

README.org

File metadata and controls

271 lines (213 loc) · 18.7 KB
Raw

RAIL Group Software Infrastructure Demos

https://github.com/RAIL-group/RAIL-software-infrastructure-demos/actions/workflows/test_unity_container.yml/badge.svg

This repository is generally devoted to demonstrating how Docker and make can be used in tandem for robotics research (and how a GPU can be supported).

This repository has a number of examples of use to Robotics and Computer Vision researchers that we make use of in the RAIL Group:

  • Run multiple processes in parallel with GNU Make. (link)
  • Run C++ code from Python using PyBind11; this code is built inside the Docker container as part of the build process. [Work in Progress]
  • Run a Unity3D simulation environment with hardware (GPU) acceleration with Virtual GL. (link)
  • Run test code inside the Docker container. (link)

Getting started

First, docker must be installed by following the official docker install guide. Once Docker works as expected, running make build from the root directory of this repository will build the Docker container used to run the demos.

Overview of Functionality

The highlights of the capabilities shown in this repository (and their associated make commands) are as follows:

  • make build A command for building the docker container used for running this repository and unzipping the Unity environment used for some of the demos
  • make all (can be run in parallel via make -j8 all) Runs tests and all demos.
  • make demo-batch-parallel A make target that aggregates other targets, each parameterized by their “random seed”. In this case, the other targets can be run in parallel, something that is easy in make using the -jN syntax, where N is the number of parallel jobs. For example, running this with make -j8 demo-batch-parallel will run 8 parallel seeds if supported by the CPU.
  • make demo-plotting Show that python plotting can be used inside the container. If the following bash environment variable is set XPASSTHROUGH=true, an interactive plot will be shown: execute export XPASSTHROUGH=true in the terminal or run the command as XPASSTHROUGH=true make demo-plot to set this variable. If this argument is not set, a plot will be written to file instead and appear in the ./data folder.
  • make demo-pybind Runs some example C++ code via PyBind11, which allows us to wrap C++ functions and classes with a Python API. This demo shows that passing an array to C++ and manipulating it via Eigen is faster than performing operations by looping over the array and competitive with Numpy.
  • make demo-unity-env Shows that a Unity environment can be run inside the Docker container, to be interfaced with via the Python unitybridge package. An example Unity has been included in unity.zip and is unpacked as part of the build process.

Other useful make targets:

  • make term Launches a bash terminal from inside the Docker container
  • make devel Like make term, but links local versions of the development code: i.e. if development code is modified inside the container, the code outside the container will also change.
  • make format Formats python code via yapf
  • make test Runs python tests via pytest

Note: this container defaults to using DISPLAY=0:0 if the DISPLAY environment variable is not set. While running the Unity environment via make demo-unity-env, you may encounter the error [VGL] ERROR: Could not open display :0; running export DISPLAY=:1 resolves this issue on most systems.

Using a GPU

This repository is configured such that if a GPU is available in the container, it can be used. For the GPU to be accessible from within the container, our docker environments will require that the NVIDIA docker runtime is installed (via nvidia-container-toolkit. Follow the install instructions on the nvidia-docker GitHub page to get it.

Once these two things are installed, you should be able to confirm that you have GPU support via:

docker run --gpus all nvidia/cuda:11.1-base nvidia-smi

Adding USE_GPU=true to any target will enable GPU-based running inside the container:

make build && make test USE_GPU=true

To confirm it is working as expected, you should notice a considerable speedup in the run time with and without the GPU for targets running the Unity simulation environment.

Running multiple processes in parallel with GNU Make

Goal: Show how to define GNU Make “targets” and execute them in parallel.

Motivation: We often generate data or run evaluations in parallel to save time. With Make, this is fairly easy, and with the right setup we can run multiple operations in parallel with only a small change at the command line.

If you are not yet familiar with the fundamentals of GNU Make, see our guide on using Make here.

Make supports arbitrary operations on strings, enabled by the `$(shell …)` syntax. This special Make command allows you to run arbitrary bash commands to generate or process strings, very useful if you would like to (for example) spawn a number of processes that have a specified (non-random) seed, and wait an amount of time between 1 and 3 seconds. We can do that with Make:

.PHONY: demo-batch-parallel-seeds demo-batch-parallel-all an-example-dependency

an-example-dependency:
	@echo "Running dependency."

demo-batch-parallel-seeds = $(shell for ii in $$(seq 100 120); do echo "demo-batch-parallel-$$ii"; done)
$(demo-batch-parallel-seeds): an-example-dependency
	@echo "Seed: $(shell echo '$@' | grep -Eo '[0-9]+'). Waiting..."
	@sleep $(shell echo '$@' | grep -Eo '[0-9]+' | awk '{print $$0%3 + 1}')
	@echo "...Done"

demo-batch-parallel-all: $(demo-batch-parallel-seeds)

So let’s break this down: first we start off with .PHONY to protect ourselves, since we’re not creating any files. Next we create an example target an-example-dependency that will serve as… an example dependency.

Next, we have a more complex series of commands that define each of our individual demo-batch-parallel-seeds targets: we first define demo-batch-parallel-seeds as a list made up of demo-batch-parallel-100 demo-batch-parallel-101 demo-batch-parallel-102 etc. For each element of the list we wish to define our make target (every item in the list is “pasted” on the left hand side of the :) and then it waits for a specified amount of time. Notice that using the built-in $(shell ...), we can do some string processing to get the number at the end of the demo-batch-parallel-## command and use it later on to control how long the sleep command runs.

Finally, the demo-batch-parallel-all target takes all of the $(demo-batch-parallel-seeds) as dependencies, which means that running `make demo-batch-parallel-all` will run every one of those other targets.

So let’s see what happens. Starting small, we can see (as we might expect) that running make an-example-dependency prints Running dependency. to the terminal. Something else interesting is that you can run a single seed individually. Calling, for example, make demo-batch-parallel-103 outputs the following (and takes roughly 2 seconds to run):

Running dependency.
Seed: 103. Waiting...
...Done

Notice that it first runs the dependency and then runs the target of interest. Now let’s see what happens when we run make demo-batch-parallel-all. The output begins with the following:

Running dependency.
Seed: 100. Waiting...
...Done
Seed: 101. Waiting...
...Done
Seed: 102. Waiting...
...Done
Seed: 103. Waiting...
...Done
Seed: 104. Waiting...
...Done
Seed: 105. Waiting...
...Done
Seed: 106. Waiting...
...Done
Seed: 107. Waiting...
...Done
Seed: 108. Waiting...
...Done

…and continues for all 20 seeds, taking a total of just over 42.5 seconds. Notice also that the dependency was only run a single time! This is what we would hope for: the dependency should only be needed a single time and Make is clever enough to have realized that, saving computation as compared to running each target individually.

Make supports parallel execution by default and makes it super easy: by adding the -j flag, followed by a number (e.g., -j4), you can run that many threads in parallel, limited only by the number of threads your processor can support. Running instead make -j4 demo-batch-parallel-all produces the following (truncated) output:

Running dependency.
Seed: 100. Waiting...
Seed: 101. Waiting...
Seed: 102. Waiting...
Seed: 103. Waiting...
...Done
Seed: 104. Waiting...
...Done
Seed: 105. Waiting...
...Done
Seed: 106. Waiting...
...Done
Seed: 107. Waiting...
...Done
Seed: 108. Waiting...
...Done
...Done
Seed: 109. Waiting...
...Done
Seed: 110. Waiting...
Seed: 111. Waiting...
...Done

The entire execution takes only 12.14 seconds, significantly faster than the original single-threaded execution, since none of the tasks block one another and can be run on separate threads.

Docker fundamentals and writing plots to file

Goal: Write a plot to file from within the Docker container; understand syntax of GNU make and how Make avoids re-generating existing files.

Motivation: We are constantly running code inside Docker and writing data or other byproducts to file. This example shows how to make that happen for a simple plotting script. Additionally, Make will save on computation when it realizes that some output already exists.

Plotting from within Docker

Make sure you have already built the repository via make build. Docker and GNU Make are at the core of our workflow. Each make target is essentially a wrapper around python. The $(DOCKER_PYTHON) variable in Make is an alias for running python inside the container. We have provided a simple plotting script and call it from a Docker container, as specified in the following Make targets:

# This target is to make an image by calling a script
demo-plotting-image-name = $(DATA_BASE_DIR)/demo_plotting.png
$(demo-plotting-image-name):
	@echo "Demo: Write a plot from within Docker"
	@$(DOCKER_PYTHON) -m scripts.plotting_demo \
		--output_image /data/demo_plotting.png

# A high-level target that calls the plotting target with a more convenient name
.PHONY: demo-plotting
demo-plotting: $(demo-plotting-image-name)

# Delete the file created by the plotting target
demo-plotting-clean:
	@echo "Cleaning products from the plotting demo."
	@echo "Are you sure? [y/N] " && read ans && [ $${ans:-N} = y ]
	@rm -rf $(demo-plotting-image-name)

Running make demo-plotting will generate an image at data/demo_plotting.png. Data created inside a Docker container is not kept by default, so we “mount” the local data in this repository at /data inside the container. When the image is written to /data/demo_plotting.png, it persists in the local folder where it can be viewed even after the container terminates.

GNU Make is clever at saving on computation. Running make demo-plotting a second time will do nothing (and Make will output Nothing to be done for `demo-plotting'. to reflect this). This is because the plot file already exists and its target (named after the file: $(DATA_BASE_DIR)/demo_plotting.png) is only run whenever that file does not exist. Delete the file by running make demo-plotting-clean. Afterwards, make demo-plotting will regenerate the file when run.

Visualizing a plot from within Docker

We also provide another target that allows one to visualize the plot without writing it to file:

.PHONY: demo-plotting-visualize
demo-plotting-visualize: XPASSTHROUGH=true
demo-plotting-visualize:
	@echo "Demo: Plotting from within Docker"
	@$(DOCKER_PYTHON) -m scripts.plotting_demo \
		--xpassthrough $(XPASSTHROUGH)

Note that this target is a bit more finicky, since it requires that the `DISPLAY` environment variable is properly set. If not, the target will fail, declaring that matplotlib is being run in `headless’ mode. Setting the display variable manually to either DISPLAY=:0 or DISPLAY=:1 will work on most machines with a working X-server:

make demo-plotting-visualize DISPLAY=:1

Running tests in Docker via PyTest

Goal: Demonstrate how to run PyTest test code from within Docker.

Motivation: Testing is an important part of any reliable workflow. Not only do manually run tests during development, but the test target is run as part of our Continuous Integration (CI) infrastructure as well. We use a GitHub Action to automatically test our code before it’s merged into main; this is also used to update the badge at the top of this repository.

Running tests is fairly straightforward:

test: build
	$(DOCKER_PYTHON) -m py.test \
		-rsx \
		--unity_exe_path /unity/$(UNITY_DBG_BASENAME).x86_64 \
		tests

Running make test will build the repository and then run the tests.

We have set up an argument to pass the Unity executable path to the tests, so that the unity environment can be run; see our conftest.py file here for details. By default-the tests are run without the GPU, but setting USE_GPU=true will enable it: make test USE_GPU=true.

Running the Unity3D environment within Docker

We have provided a script and accompanying make target that runs the Unity simulation environment:

# A target that runs the Unity3D enviornment and generates an image
.PHONY: demo-unity-env
demo-unity-env:
	  @echo "Demo: Interfacing with Unity"
	  @$(call xhost-activate)
	  @docker run --init --net=host \
		  $(DOCKER_ARGS) $(DOCKER_CORE_ARGS) \
		  ${IMAGE_NAME}:${VERSION} \
		  python3 -m scripts.unity_env_demo \
		  --unity_exe_path /unity/$(UNITY_DBG_BASENAME).x86_64 \
		  --output_image /data/demo_unity_env.png \
		  --xpassthrough $(XPASSTHROUGH)

Now you can run this code using one of these configurations:

# With the CPU
make build && make demo-unity-env
# With a GPU
make build && make demo-unity-env USE_GPU=true
# With a GPU (some machines use DISPLAY=:1 and will fail without this)
make build && make demo-unity-env USE_GPU=true DISPLAY=:1

Running one of these should write an image demo_unity_env.png into the data folder. With USE_GPU=true, the Unity3D environment runs with hardware acceleration (as long as Nvidia Docker is configured and runs with your local GPU), allowing us to generate the image relatively quickly. Note: There may be a number of warnings beginning with ALSA upon running this command. These are complaints that a sound card does not exist and can be ignored for our purposes.

The Unity environment can also be run on the CPU (the configuration above without USE_GPU=true) though is considerably slower. However, this feature can be useful for running simple unit tests, and indeed a test confirming that we can communicate with the Unity simulation environment is included in our unit tests and is run as part of our automated continuous integration setup managed via GitHub Actions.

How does hardware acceleration (GPU) work?

The Unity3D environment is running inside a Docker container with hardware support. There are a few pieces required to make this setup work correctly. The first is the container itself, which must have the ability to support OpenGL. For this, our Dockerfile starts with the cudagl container provided by Nvidia:

FROM nvidia/cudagl:11.1-devel-ubuntu20.04

This container has OpenGL already installed and provides the resources we need to access hardware acceleration that Unity3D relies upon to run at target speeds. Next, we need to build VirtualGL inside the container; VirtualGL was created to allow for “server-side 3D rendering” where a computer may not have a screen attached and may or may not be running an X window server. VirtualGL allows to use local hardware (a GPU) to run applications on this “remote machine” (or inside a container). To build VirtualGL, we use the following command in the Dockerfile:

# Install VirtualGL
ENV VIRTUALGL_VERSION 2.5.2
RUN curl -sSL https://downloads.sourceforge.net/project/virtualgl/"${VIRTUALGL_VERSION}"/virtualgl_"${VIRTUALGL_VERSION}"_amd64.deb -o virtualgl_"${VIRTUALGL_VERSION}"_amd64.deb && \
	    dpkg -i virtualgl_*_amd64.deb && \
	    /opt/VirtualGL/bin/vglserver_config -config +s +f -t && \

Finally, we need to run our Unity environment. VirtualGL still requires an X window server to be running, so we “fake” one using xvfb (the X virtual frame buffer), which creates a X window server inside the docker container that VirtualGL can latch on to. The full code exists inside the src/entrypoint.sh script that launches when the Docker container is created, but the relevant snippet is here:

export VGL_DISPLAY=$DISPLAY
xvfb-run -a --server-num=$((99 + $RANDOM % 10000)) \
     --server-args='-screen 0 640x480x24 +extension GLX +render -noreset' vglrun $@

This script does a number of things all at once:

  1. It sets VGL_DISPLAY to $DISPLAY, which is required to ensure that the “display” that VirtualGL is writing to is the same as the display the X window manager is writing to. Even without a physical display, this is important for GPU access.
  2. It launches a local X window server using xvfb-run. The --server-num is set to a random number so that there is no conflict during container creation; even though there could be a conflict, xvfb will find another number if one is already running. This is helpful to avoid a race condition, that should no longer happen (…much), deconflicted due to $RANDOM. The arguments on the right tell xvfb to create a screen of size 640x480 and with GLX: OpenGL hardware acceleration.
  3. Finally, vglrun $@ runs the command input to the script $@ within VirtualGL (and to inherit the X window server provided by xvfb-run.

Alltogether, this allows us to run any process inside the Docker container with hardware-acceleration and an X window server.

We provide the unitybridge package that launches our Unity environment, which is provided as a pre-built binary along with this repository. Unfortunately, at this time the Unity environment itself is not open source, but this process should work with any Unity application.