You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We need to complete the initial draft https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/README.md.
1
+
# Building PaddlePaddle
2
2
3
-
I am recording some ideas here, and we should file a PR later.
3
+
## Goals
4
4
5
-
## Current Status
5
+
We want the building procedure generates Docker images, so we can run PaddlePaddle applications on Kubernetes clusters.
6
6
7
-
Currently, we have four sets of Dockefiles:
7
+
We want it generates .deb packages, so that enterprises without Docker support can run PaddlePaddle applications as well.
8
8
9
-
1. Kubernetes examples:
9
+
We want to minimize the size of generated Docker images and .deb packages so to ease the deployment cost.
10
10
11
-
```
12
-
doc/howto/usage/k8s/src/Dockerfile -- based on released image but add start.sh
13
-
doc/howto/usage/k8s/src/k8s_data/Dockerfile -- contains only get_data.sh
14
-
doc/howto/usage/k8s/src/k8s_train/Dockerfile -- this duplicates with the first one.
15
-
```
11
+
We want to encapsulate building tools and dependencies in a *development* Docker image so to ease the tools installation for developers.
16
12
17
-
1. Generate .deb packages:
13
+
We want developers can use whatever editing tools (emacs, vim, Eclipse, Jupyter Notebook), so the development Docker image contains only building tools, not editing tools, and developers are supposed to git clone source code into their development computers, instead of the container running the development Docker image.
18
14
19
-
```
20
-
paddle/scripts/deb/build_scripts/Dockerfile -- significantly overlaps with the `docker` directory
21
-
```
15
+
We want the procedure and tools work also with testing, continuous integration, and releasing.
22
16
23
-
1. In the `docker` directory:
24
17
25
-
```
26
-
paddle/scripts/docker/Dockerfile
27
-
paddle/scripts/docker/Dockerfile.gpu
28
-
```
29
-
30
-
1. Document building
31
-
32
-
```
33
-
paddle/scripts/tools/build_docs/Dockerfile -- a subset of above two sets.
34
-
```
35
-
36
-
## Goal
18
+
## Docker Images
37
19
38
20
We want two Docker images for each version of PaddlePaddle:
39
21
@@ -45,7 +27,9 @@ We want two Docker images for each version of PaddlePaddle:
45
27
- release engineers -- use this to build the official release from certain branch/tag on Github.com.
46
28
- document writers / Website developers -- Our documents are in the source repo in the form of .md/.rst files and comments in source code. We need tools to extract the information, typeset, and generate Web pages.
47
29
48
-
So the development image must contain not only source code building tools, but also documentation tools:
30
+
Of course developers can install building tools on their development computers. But different version of PaddlePaddle might require different set/version of building tools. Also, it makes collaborative debugging eaiser if all developers use a unified development environment.
31
+
32
+
The development image should include the following tools:
49
33
50
34
- gcc/clang
51
35
- nvcc
@@ -54,7 +38,7 @@ We want two Docker images for each version of PaddlePaddle:
54
38
- woboq
55
39
- sshd
56
40
57
-
where `sshd` makes it easy for developers to have multiple terminals connecting into the container.
41
+
where `sshd` makes it easy for developers to have multiple terminals connecting into the container.`docker exec` works too, but if the container is running on a remote machine, it would be easier to ssh directly into the container than ssh to the box and run `docker exec`.
58
42
59
43
1.`paddle:<version>`
60
44
@@ -65,82 +49,107 @@ We want two Docker images for each version of PaddlePaddle:
65
49
- no-GPU/AVX `paddle:<version>`
66
50
- no-GPU/no-AVX `paddle:<version>-noavx`
67
51
68
-
We'd like to give users choices of GPU and no-GPU, because the GPU version image is much larger than then the no-GPU version.
52
+
We'd like to give users the choice between GPU and no-GPU, because the GPU version image is much larger than then the no-GPU version.
53
+
54
+
We'd like to give users the choice between AVX and no-AVX, because some cloud providers don't provide AVX-enabled VMs.
55
+
56
+
57
+
## Development Environment
58
+
59
+
Here we describe how to use above two images. We start from considering our daily development environment.
60
+
61
+
Developers work on a computer, which is usually a laptop or desktop:
62
+
63
+
64
+
65
+
or, they might rely on a more sophisticated box (like with GPUs):
66
+
67
+
68
+
69
+
A basic principle is that source code lies on the development computer (host), so that editing tools like Eclipse can parse the source code and support auto-completion.
70
+
69
71
70
-
We'd like to give users choices of AVX and no-AVX, because some cloud providers don't provide AVX-enabled VMs.
72
+
## Usages
71
73
72
-
##Dockerfile
74
+
### Build the Development Docker Image
73
75
74
-
To realize above goals, we need only one Dockerfile for the development image. We can put it in the root source directory.
76
+
The following commands check out the source code on the development computer (host) and build the development image `paddle:dev`:
75
77
76
-
Let us go over our daily development procedure to show how developers can use this file.
The `docker build` command assumes that `Dockerfile` is in the root source tree. This is reasonable because this Dockerfile is this only on in our repo in this design.
docker build -t paddle:dev .# Suppose that the Dockerfile is in the root source directory.
96
-
```
95
+
This command mounts the source directory on the host into `/paddle` in the container, so the default entrypoint of `paddle:dev`, `build.sh`, would build the source code with possible local changes. When it writes to `/paddle/build` in the container, it actually writes to `$PWD/build` on the host.
-`$PWD/build/paddle-<version>.deb` for production installation, and
101
+
-`$PWD/build/Dockerfile`, which builds the production Docker image.
103
102
104
-
This command maps the source directory on the host into `/paddle` in the container.
105
103
106
-
Please be aware that the default entrypoint of `paddle:dev` is a shell script file `build.sh`, which builds the source code, and outputs to `/paddle/build` in the container, which is actually `$PWD/build` on the host.
104
+
### Build the Production Docker Image
107
105
108
-
`build.sh` doesn't only build binaries, but also generates a `$PWD/build/Dockerfile` file, which can be used to build the production image. We will talk about it later.
106
+
The following command builds the production image:
109
107
110
-
1. Run on the host (Not recommended)
108
+
```bash
109
+
docker build -t paddle -f build/Dockerfile .
110
+
```
111
111
112
-
If the host computer happens to have all dependent libraries and Python runtimes installed, we can now run/test the built program. But the recommended way is to running in a production image.
112
+
This production image is minimal -- it includes binary `paddle`, the share library `libpaddle.so`, and Python runtime.
113
113
114
-
1. Run in the development container
114
+
###Run PaddlePaddle Applications
115
115
116
-
`build.sh` generates binary files and invokes `make install`. So we can run the built program within the development container. This is convenient for developers.
116
+
Again the development happens on the host. Suppoose that we have a simple application program in `a.py`, we can test and run it using the production image:
117
117
118
-
1. Build a production image
118
+
```bash
119
+
docker run -it -v $PWD:/work paddle /work/a.py
120
+
```
119
121
120
-
On the host, we can use the `$PWD/build/Dockerfile`to generate a production image.
122
+
But this works only if all dependencies of `a.py` are in the production image. If this is not the case, we need to build a new Docker image from the production image and with more dependencies installs.
We need a Dockerfile in https://github.com/paddlepaddle/book that builds Docker image `paddlepaddle/book:<version>`, basing on the PaddlePaddle production image:
127
127
128
-
Once we have the production image, we can run [Paddle Book](http://book.paddlepaddle.org/) chapters in Jupyter Notebooks (if we chose to build them)
128
+
```
129
+
FROM paddlepaddle/paddle:<version>
130
+
RUN pip install -U matplotlib jupyter ...
131
+
COPY . /book
132
+
EXPOSE 8080
133
+
CMD ["jupyter"]
134
+
```
129
135
130
-
```bash
131
-
docker run -it paddle
132
-
```
136
+
The book image is an example of PaddlePaddle application image. We can build it
133
137
134
-
Note that the default entrypoint of the production image starts Jupyter server, if we chose to build Paddle Book.
138
+
```bash
139
+
git clone https://github.com/paddlepaddle/book
140
+
cd book
141
+
docker build -t book .
142
+
```
135
143
136
-
1.Run on Kubernetes
144
+
### Build and Run Distributed Applications
137
145
138
-
We can push the production image to a DockerHub server, so developers can run distributed training jobs on the Kuberentes cluster:
146
+
In our [API design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/api.md#distributed-training), we proposed an API that starts a distributed training job on a cluster. This API need to build a PaddlePaddle application into a Docekr image as above, and calls kubectl to run it on the cluster. This API might need to generate a Dockerfile look like above and call `docker build`.
139
147
140
-
```bash
141
-
docker tag paddle me/paddle
142
-
docker push
143
-
kubectl ...
144
-
```
148
+
Of course, we can manually build an application image and launch the job using the kubectl tool:
145
149
146
-
For end users, we will provide more convinient tools to run distributed jobs.
0 commit comments