Reduce docker image size for runtime usages about autoware

[Sharrrrk]

Currently I have started to change Dockerfile and ansible scripts, the main ideas:

• Use cuda runtime images as base
• Only runtime packages of TensorRT are installed
• Only minimal set of ROS packages are installed
• Multistage image build, copy install folder

[xmfcx]

I've created this discussion on re-layering of the dockerfile/image:

https://github.com/orgs/autowarefoundation/discussions/3469

[kaspermeck-arm]

Merge this issue with autowarefoundation/autoware.universe#4111.

[Sharrrrk]

With docker image built from nvidia/cuda:11.4.0-runtime-ubuntu20.04, and runtime dependencies installed for autoware, the disk usage of /usr/local/cuda-11.4 folder could be reduced by 2.2GB.

Original: 4.0G /usr/local/cuda-11.4
New: 1.8G /usr/local/cuda-11.4

DNN related runtime dependencies:

• libcudnn8
• libnvinfer8
• libnvinfer-plugin8
• libnvonnxparsers8
• libnvparsers8

[Sharrrrk]

Currently I'm thinking of two approaches of this work:

• Reduce size of prebuilt image
  Considering the prebuilt images is for users to try out Autoware quickly, dev environment could be remove from it.

• Keep current prebuilt image and build another runtime image
  This allows users to rebuild partial of Autoware, not the whole project using prebuilt image, then runtime image may be provided just for execution.

The main concept of size reducing:

• Only use runtime libraries for CUDA/TensorRT (This could save ~2.2GB)
• Only copy the install folder or maybe also keep src folder (This could save up to ~3.3GB)

@kenji-miyake, what's your opinion about it ?

[kenji-miyake]

@Sharrrrk Thank you for investigating this issue!
I totally agree with you, and either option seems okay. The first option would be better as a first step because it's simpler.

[kaspermeck-arm]

@Sharrrrk thanks for taking this on!

For inspiration, you could take a look at the images layers which ROS and NVIDIA have used to create their Docker images. See

• https://github.com/osrf/docker_images/blob/3f4fbca923d80f834f3a89b5960bad5582652519/ros/galactic/ubuntu/focal/ros-core/Dockerfile
• https://hub.docker.com/layers/nvidia/cuda/11.4.2-runtime-ubuntu18.04/images/sha256-9fd4770341bd8f2cd754b70d552dfd0cc5e6e3ed1406449f00b1c5bb944133b0?context=explore

I'd recommend building up a new container and not removing from an existing one.

[Sharrrrk]

Hi @kasperornmeck, I looked into the docker file of Nvidia and ROS, Nvidia used a script from outside to install all the packages, and for ROS, it's just like what have been implented in current ansible script. So I would say that use nvidia runtime docker image as base would be a good and neat approach. I will modify the ansible srcipts (add a variant that only installs runtime dependencies) and the Docker file (use mutlti stage build) of course, so that the prebuilt image could reduce it's size in later build.

[Sharrrrk]

Hi @kasperornmeck, I looked into the docker file of Nvidia and ROS, Nvidia used a script from outside to install all the packages, and for ROS, it's just like what have been implented in current ansible script. So I would say that use nvidia runtime docker image as base would be a good and neat approach. I will modify the ansible srcipts (add a variant that only installs runtime dependencies) and the Docker file (use mutlti stage build) of course, so that the prebuilt image could reduce it's size in later build.

[kaspermeck-arm]

Hi @Sharrrrk
I was installing the NVIDIA driver using the runfile (local) method. You might already be aware of this but you can install only the driver, or any of the other software by selecting/deselecting, e.g., all the runtime dependencies.

lqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqk
x CUDA Installer                                                               x
x - [X] Driver                                                                 x
x      [X] 515.65.01                                                           x
x - [ ] CUDA Toolkit 11.7                                                      x
x    - [ ] CUDA Libraries 11.7                                                 x
x       + [ ] CUDA Development 11.7                                            x
x       - [ ] CUDA Runtime 11.7                                                x
x            [ ] cuda-cudart                                                   x
x            [ ] cuda-nvrtc                                                    x
x            [ ] libcublas11                                                   x
x            [ ] libcufft                                                      x
x            [ ] libcurand                                                     x
x            [ ] libcusolver                                                   x
x            [ ] libcusparse                                                   x
x            [ ] libnpp                                                        x
x            [ ] libnvjpeg                                                     x
x    - [ ] CUDA Tools 11.7                                                     x
x       + [ ] CUDA Command Line Tools 11.7                                     x
x       + [ ] CUDA Visual Tools 11.7                                           x
x    - [ ] CUDA Compiler 11.7                                                  x
x         [ ] cuda-cuobjdump                                                   x
x                                                                              x
x Up/Down: Move | Left/Right: Expand | 'Enter': Select | 'A': Advanced options x

[Sharrrrk]

Hi @kasperornmeck thanks for the information, I will compare the docker size between manually install and cuda-runtime.

[BonoloAWF]

@Sharrrrk have you created a PR for these changes? If so, please link it to the Issue.

[xmfcx]

Closing this issue since it will be addressed by:

• #3999