By default, Ubuntu or Debian based "apt" or "apt-get" system installs recommended but not suggested packages .
By passing "--no-install-recommends" option, the user lets apt-get know not to consider recommended packages as a dependency to install.
This results in smaller downloads and installation of packages .
Refer to blog at [Ubuntu Blog](https://ubuntu.com/blog/we-reduced-our-docker-images-by-60-with-no-install-recommends) .
Signed-off-by: Pratik Raj <rajpratik71@gmail.com>
xref #6728
When install the nvidia docker runtime, kubelet will try to restart
stopped containers via old docker. This will leak those containers as
orphan containers.
* Changed Dockerfile base image to debian for systemctl and bash.
* Added autodetect of AWS ec2 instanceclass p2, p3, g3.
* For each detected instance class, added the installation of the proper driver
according to the specific NVIDIA hardware.
- G3 instance types require Nvidia Grid Series/Grid K520 drivers
- P2 instance types require Nvidia Tesla K-Series drivers
- P3 instance types require Nvidia Tesla V-Series drivers
* Set custom nvidia-smi configurations according to nvidia hardware per ec2
instanceclass, according to the AWS GPU optimization document.
* Added the installation and patches of the latest cuda 9.1 libraries.
* Added restart of kubelet on kube node at end of successful hook run, thereby
fixing a race condition where kubelet would start before the Nvidia drivers
were loaded, thus not allowing kubernetes to detect GPUS on the kube node.
* Ensured build of nvidia drivers used same gcc version as that which built
default kops kernel.
* Fixed issue where *every* run of this container would download all the NVIDIA
drivers + cuda libs (1GB+), by caching the files on the kube node.
* Fixed issue where after reboot, subsequent runs of this script would fail
because mknod would try to create a previously-created device node and fail.
This previously caused download loop as systemd perpetually restarted the
unit upon failure.
* Tested with p2.xlarge, p3.2xlarge, and g3.4xlarge
Ciprian Hacman
Pratik raj
Huihua Zhang
Hanfei Shen
tanjunchen
Adrian Lyjak
Robert Everson
k8s-ci-robot
Steve Winslow
David C Wang
Justin Santa Barbara