README.md 5.5 KB
Newer Older
1
# Docker Builds for the NGraph-Cpp _Reference-OS_
2

3
## Introduction
4

5
This directory contains a basic build system for creating docker images of the _reference-OS_ on which NGraph-Cpp builds and unit tests are run.  The purpose is to provide reference builds for _Continuous Integration_ used in developing and testing NGraph-Cpp.
6

7
The `Makefile` provides targets for:
8

9 10 11 12 13
* Building the _reference-OS_ into a docker image
* Building NGraph-Cpp and running unit tests in this cloned repo, mounted into the docker image of the _reference-OS_
* Starting an interactive shell in the _reference-OS_ docker image, with the cloned repo available for manual builds and unit testing

The _make targets_ are designed to handle all aspects of building the _reference-OS_ docker image, running NGraph-Cpp builds and unit testing in it, and opening up a session in the docker image for interactive use.  You should not need to issue any manual commands (unless you want to).  In addition the `Dockerfile.ngraph_cpp_cpu` provides a description of how the _reference-OS_ is built, should you want to build your own server or docker image.
14

15
## Prerequisites
16

17
In order to use the make targets, you will need to do the following:
18

19 20 21 22
* Have docker installed on your computer with the docker daemon running.
* These scripts assume that you are able to run the `docker` command without using `sudo`.  You will need to add your account to the `docker` group so this is possible.
* If your computer (running docker) sits behind a firewall, you will need to have the docker daemon properly configured to use proxies to get through the firewall, so that public docker registries and git repositories can be accessed.
* You should _not_ run `make check_cpu` from a directory in an NFS filesystem, if that NFS filesystem uses _root squash_ (see **Notes** section below).  Instead, run `make check_cpu` from a cloned repo in a local filesystem.
23

24
## Make Targets
25

26 27
The _make targets_ are designed to provide easy commands to run actions using the docker image.  All _make targets_ should be issued on the host OS, and _not_ in a docker image.

28
* In general, you simply need to run the command **`make check_cpu`**.  This first makes the `build_ngraph_cpp_cpu` target as a dependency.  Then it makes the `check_cpu` target, which will build NGraph-Cpp using _cmake_ and _make_ and then run unit testing.  Please keep in mind that `make check_cpu` does not work when your working directory is in an NFS filesystem that uses _root squash_ (see **Notes** section below).
29

30
* You can also run the command **`make shell`** to start an interactive bash shell inside the docker image.  While this is not required for normal builds and unit testing, it allows you to run interactively within the docker image with the cloned repo mounted.  Again, `build_ngraph_cpp_cpu` is made first as a dependency.  Please keep in mind that `make shell` does not work when your working directory is in an NFS filesystem that uses _root squash_ (see **Notes** section below).
31 32 33

* Running the command **`make build_ngraph_cpp_cpu`** is also available, if you simply want to build the docker image.  This target does work properly when your working directory is in an NFS filesystem.

34
Note that all operations performed inside the the docker image are run as **root**.  This has unfortunate side effects if you run a make target that does operations inside the docker image, while your working directory is in an NFS filesystem that uses _root squash_ (see **Notes** below).
35 36 37

## Helper Scripts

38
These helper scripts are included for use in the `Makefile` and automated (Jenkins) jobs.  **These scripts should _not_ be called directly unless you understand what they do.**
39

40
#### `chown_files.sh`
41 42 43

Used in the `Makefile` to change the ownership of files while running _inside the docker image_.  This is used to fix a problem with docker where files written in a mounted filesystem, from within the docker image, will end up being owned by root in the host OS.  This leads to problems with Jenkins not being able to clean up its own job directories if docker images are used.

44
#### `docker_cleanup.sh`
45

46
A helper script for Jenkins jobs to clean up old exited docker containers and `ngraph_*` docker images.
47 48 49

## Notes

50
* The top-level `Makefile` in this cloned repo can be used to build and unit-test NGraph-Cpp _outside_ of docker.  This directory is only for building and running unit tests for NGraph-Cpp in the _reference-OS_ docker image.
51

52
* The `_cpu` suffix in the targets refers to the docker image being built such that testing is in a _cpu-only_ environment.  At a later date, _gpu_ and other environments may be added.  This convention was chosen to stay consistent with the _NervanaSystems ngraph_ project on Github.
53

54 55
* Due to limitations in how docker mounts work, `make check_cpu` and `make shell` will fail
if you try to run them while in a working directory that is in an NFS-mount that has _root squash_ enabled.  The cause results from the process in the docker container running as root.  When a file or directory is created by root in the mounted directory tree, from within the docker image, the NFS-mount (in the host OS) does not allow a root-created file, leading to a permissions error.  This is dependent on whether the host OS performs "root squash" when mounting NFS filesystems.  The fix to this is easy: run `make check_cpu` and `make shell` from a local filesystem.
56 57

* These make targets have been tested on Ubuntu 14.04 and Ubuntu 16.04 with docker installed and the docker daemon properly configured.  Some adjustments may be needed to run these on other OSes.