Thinner docker image #1598

typhoonzero · 2017-03-10T09:13:22Z

build a docker image to build paddle
copy out binaries
build the paddle-core image

TODO:
other images like jupyter and woboq stuff?

gangliao · 2017-03-10T09:49:59Z

This figure shows that the size of new docker image based on this PR is far less than the current image "0.10.0rc1-cpu".

Yancey1989 · 2017-03-10T09:47:25Z

paddle/scripts/docker/buildall.sh

+
+function build_in_docker() {
+  docker build . -t paddle-build-env -f paddle/scripts/docker/buildimage/Dockerfile.build
+  BUILDER=$(docker run -d paddle-build-env)


这里用docker run -d --rm --name paddle-builder paddle-build-env吧，也不需要$BUILDER和后面的docker rm 了

helinwang · 2017-03-10T19:13:37Z

赞👍，有办法能让这个方式（docker build两次）在dockerhub跑起来吗？如果不行的话我来把我们的CI服务器设置一下专门搞这个吧。感觉瘦身版的docker image非常有必要。

wangkuiyi · 2017-03-12T06:32:08Z

paddle/scripts/docker/buildall.sh

+  BUILDER=$(docker run -d paddle-build-env)
+  docker exec $BUILDER /bin/bash -c "export BUILD_AND_INSTALL=ON && /paddle/paddle/scripts/docker/build.sh"
+  mkdir -p $BINARIES_DIR
+  docker cp $BUILDER:/usr/local/opt/paddle/bin/paddle_pserver_main $BINARIES_DIR


这个PR看上去和我在 #1599 (comment) 描述的还是不一样。我那个方法不需要 docker cp，因为源码目录 paddle以及其中的子目录paddle/build都在host上，是通过 docker run -v $PWD:/paddle paddlepaddle/paddle:dev 中的 -v参数分享给 development container 的，所以二进制文件自然就在 paddle/build 子目录里了。

gongweibao · 2017-03-12T14:17:34Z

转到#1602

reyoung · 2017-03-13T02:02:01Z

paddle/scripts/docker/buildall.sh

+  docker cp $BUILDER:/usr/local/bin/paddle $BINARIES_DIR
+  docker cp $BUILDER:/usr/local/opt/paddle/bin/paddle_usage $BINARIES_DIR
+
+  docker cp $BUILDER:/usr/local/opt/paddle/share/wheels $BINARIES_DIR


我记得Paddle的CMake是可以pack出来deb包的。如果想要编译出来一个可以被其他镜像用的二进制包，不如直接把编译的时候打包成deb，再把deb拷贝出来。这样做的好处是:

不用写死拷贝的路径了。

修改CMake文件来增加新二进制文件的时候，不用修改dockerfile了。

打包deb的代码好久没动了，具体的可以参考这里

reyoung · 2017-03-13T02:06:29Z

paddle/scripts/docker/buildimage/Dockerfile.build

+
+ENV HOME /root
+
+RUN sed 's@http:\/\/archive.ubuntu.com\/ubuntu\/@mirror:\/\/mirrors.ubuntu.com\/mirrors.txt@' -i /etc/apt/sources.list && \


替换成 mirror://mirrors.ubuntu.com/mirrors.txt，会导致ubuntu的一些PPA报 hashsum mismatched 错误。主要原因是，有一些mirror的同步周期比较长，导致PPA里面依靠的软件包还没有出现在mirror里面。

感觉默认情况下用ubuntu的官方源是靠谱的。之前有一个编译参数是 UBUNTU_MIRROR。如果需要使用mirror，就在编译Paddle的时候，添加编译参数即可(参考这个文件的第六行第七行)。

docker build --build-arg UBUNTU_MIRROR=mirror://mirrors.ubuntu.com/mirrors.txt

我测试的情况看，造成hashsum mismatched 错误的原因应该是GFW的问题。添加新的mirrors之后，这种情况基本不再出现。同时，update的时候应该会检测最近的源在哪里。所以还是建议把mirrors加到里边。

我的建议也是在代码中使用Ubuntu的官方源，可以在文档中说明替换源的方法。

wangkuiyi

To my understand, if we are going to implement the design, we shouldn't add any more Dockefiles in this PR. Instead, we should have only one Dockerfile.dev, which builds the development environment.

wangkuiyi · 2017-03-15T05:51:51Z

paddle/scripts/docker/buildall.sh

+}
+
+function build_paddle_core() {
+  docker build . -t paddle-core:$VERSION -f paddle/scripts/docker/paddle-core/Dockerfile


@reyoung and I worked hard and reduced the number of Docker images. Why we'd need this many Docker images here? That is core for? This differs from the two images as described in our design doc.

Will rename "paddle-core" to paddle. It is the production image

wangkuiyi · 2017-03-15T22:54:34Z

.gitignore

@@ -1,5 +1,7 @@
 *.DS_Store
-build/
+build/paddle/math


git必须忽略整个build目录吧。

.dockerignore 也应该忽略整个build目录，因为实际上使用的是host上的build目录。

Done.

改错了。现在使用deb包安装，所以可以忽略build目录了。

wangkuiyi · 2017-03-15T22:55:38Z

paddle/scripts/deb/build_scripts/build.sh

 cd build
-cmake .. -DWITH_GPU=OFF -DWITH_SWIG_PY=ON -DWITH_AVX=ON
+cmake .. -DWITH_GPU=OFF -DWITH_SWIG_PY=ON -DWITH_AVX=ON -DWITH_SWIG_PY=ON -DWITH_STYLE_CHECK=OFF -DCMAKE_EXPORT_COMPILE_COMMANDS=ON


这里这些变量都写死了，那如果要支持 GPU怎么办呢？

这个脚本是会用cpack打包4个deb包，分别对应cpu, cpu-noavx, gpu, gpu-noavx，在制作dockerimage的时候决定使用哪个

另外，后文说到要去掉 paddle/scripts/deb 目录。那这个文件是不是就不需要了？

wangkuiyi · 2017-03-15T22:56:13Z

paddle/scripts/docker/README.md

@@ -0,0 +1,9 @@
+# Build docker image
+
+We use a docker environment to build paddle binaries and put it into a runtime image `paddle-core` for uses of most cases


build的结果是不是就放在host的文件系统里就行了？

目前已经改成这样了，是mount host的目录来build的

… thinnerdocker

typhoonzero · 2017-03-16T03:08:53Z

Will continue work to fit #1625 and #1627

helinwang · 2017-03-16T21:05:10Z

paddle/scripts/deb/build_scripts/build.sh

-mkdir build
+
+# clean build dir and third_party dir cache
+rm -rf build third_party


不删除build文件夹有坏处吗？好处应该可以加速build过程。好像make可以自动追踪什么文件被改变了，感觉不会有问题？
要是用户想要clean build可以执行build.sh之前rm -rf build或者有个参数./build.sh clean

之前编译遇到不删除build和thirdparty，导致protobuf的依赖没被cmake找到。或许可以只删除CmakeCache?

明白了，没事，这个不影响merge。

helinwang · 2017-03-16T21:07:49Z

paddle/scripts/deb/build_scripts/build.sh

 make -j `nproc`
+# FIXME(typhoonzero): add ARCH gpu noavx flag to CPACK_SYSTEM_NAME. Why -D not affect anything?
 cpack -D CPACK_GENERATOR='DEB' ..
 mv *.deb ~/dist/cpu

 rm -rf *


不是很了解cpack怎么工作的，这里和以后的几处rm -rf *需要吗？。cmake好像不会受干扰，而且不删掉反而能加速编译过程？

根据 #1625 ，考虑直接去掉paddle/scripts/deb目录了，使用docker目录下的build.sh完成

helinwang · 2017-03-16T21:30:51Z

paddle/scripts/docker/buildall.sh

+  # FIXME: need to wait a signal not sleeping
+  BUILDER=$(docker run -d -v ${PWD}:/root/paddle -v ${DEB_DIST_DIR}:/root/dist paddle-build-env sleep 3600)
+  # NOTICE: build deb files for real paddle image
+  docker exec $BUILDER /bin/bash -c "/root/paddle/paddle/scripts/deb/build_scripts/build.sh"


以上两句是不是可以改成

docker run --rm -v ${PWD}:/root/paddle -v ${DEB_DIST_DIR}:/root/dist paddle-build-env bash /root/paddle/paddle/scripts/deb/build_scripts/build.sh

然后命令结束的时候对应的返回码应该也是正确的（docker run ... ; echo $?），
接下来的docker stop $BUILDER && docker rm $BUILDER也可以删掉了。

根据 #1625 去掉了buildall.sh 自动build image的过程新开一个PR，使用travis ci完成

gangliao · 2017-03-17T06:45:38Z

paddle/scripts/docker/build.sh

+  MIRROR_UPDATE="\\"
+fi
+
+cat > /paddle/build/Dockerfile <<EOF


/paddle/build/Dockerfile 名字是不是也得模板化啊

与多个tags对应

是指加上.gpu,. noavx的后缀？

对啊

from @wangkuiyi

我们应该会有一个build.sh，它通过执行一个Dockerfile template来生成对应不同tag的Dockerfiles。我理解最好这个模板本身就在build.sh里，而不是多个tags对应的Dockerfiles在git repo里。

gangliao

LGTM

wangkuiyi · 2017-03-17T15:28:21Z

Dockerfile

@@ -14,14 +16,12 @@ ARG WITH_STYLE_CHECK

 ENV BUILD_WOBOQ=${BUILD_WOBOQ:-OFF}
 ENV BUILD_AND_INSTALL=${BUILD_AND_INSTALL:-OFF}
-ENV WITH_GPU=ON
+ENV WITH_GPU=OFF


这里应该加一个

ARG WITH_GPU ENV WITH_GPU=${WITH_GPU:-OFF}

wangkuiyi · 2017-03-17T15:30:19Z

paddle/scripts/deb/build_scripts/build.sh

@@ -1,5 +1,6 @@
 #!/bin/bash
 set -e
+ARCH=$(uname -i)


什么时候会调用这个脚本呢？我看Dockerfile里调用的是 paddle/scripts/docker/build.sh 。

考虑在别的PR中删除"deb"这个目录了 @helinwang

helinwang

LGTM. 不影响merge：paddle/scripts/deb/build_scripts/build.sh这个可以删掉了吗？

Yancey1989

LGTM

thinner docker

a6eec02

typhoonzero requested review from Yancey1989 and gangliao March 10, 2017 09:13

gangliao assigned helinwang Mar 10, 2017

Yancey1989 reviewed Mar 10, 2017

View reviewed changes

gangliao assigned gangliao and helinwang and unassigned helinwang Mar 10, 2017

typhoonzero mentioned this pull request Mar 10, 2017

modify dockerfile to autobuild on dockerhub #1599

Closed

wangkuiyi requested changes Mar 12, 2017

View reviewed changes

gongweibao mentioned this pull request Mar 12, 2017

Build docker image #1602

Merged

gongweibao closed this Mar 12, 2017

gangliao reopened this Mar 12, 2017

reyoung reviewed Mar 13, 2017

View reviewed changes

yi.wu added 2 commits March 13, 2017 19:00

do not use docker cp

f66fd44

use deb to build paddle-core image

1b269db

typhoonzero changed the title ~~[WIP] Thinner docker image~~ Thinner docker image Mar 14, 2017

wangkuiyi requested changes Mar 15, 2017

View reviewed changes

wangkuiyi reviewed Mar 15, 2017

View reviewed changes

yi.wu added 3 commits March 16, 2017 09:55

update .gitignore

12a3af1

production image name paddle not paddle-core

3298d30

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

c1d5aaa

… thinnerdocker

helinwang reviewed Mar 16, 2017

View reviewed changes

new design: PaddlePaddle#1625

1ab419a

gangliao mentioned this pull request Mar 17, 2017

Improve Docker Images #1630

Closed

5 tasks

gangliao reviewed Mar 17, 2017

View reviewed changes

tags for build docker image

653c981

gangliao approved these changes Mar 17, 2017

View reviewed changes

yi.wu added 2 commits March 17, 2017 20:55

add python requests, add DELETE_BUILD_CACHE arg

0b4d456

gpu image dependencies

e8d33be

wangkuiyi approved these changes Mar 17, 2017

View reviewed changes

helinwang approved these changes Mar 17, 2017

View reviewed changes

Yancey1989 approved these changes Mar 18, 2017

View reviewed changes

gongweibao merged commit 21e878f into PaddlePaddle:develop Mar 18, 2017

typhoonzero mentioned this pull request Mar 18, 2017

Add GPU arg in develop Dockerfile #1639

Merged

typhoonzero deleted the thinnerdocker branch August 11, 2017 06:54


		ENV HOME /root

		RUN sed 's@http:\/\/archive.ubuntu.com\/ubuntu\/@mirror:\/\/mirrors.ubuntu.com\/mirrors.txt@' -i /etc/apt/sources.list && \

		@@ -0,0 +1,9 @@
		# Build docker image

		We use a docker environment to build paddle binaries and put it into a runtime image `paddle-core` for uses of most cases

Thinner docker image #1598

Thinner docker image #1598

Conversation

typhoonzero commented Mar 10, 2017 • edited by wangkuiyi Loading

gangliao commented Mar 10, 2017 • edited Loading

Choose a reason for hiding this comment

helinwang commented Mar 10, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gongweibao commented Mar 12, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gongweibao Mar 13, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

wangkuiyi left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

typhoonzero Mar 16, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

typhoonzero commented Mar 16, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

helinwang Mar 16, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

helinwang Mar 16, 2017 • edited Loading

Choose a reason for hiding this comment

typhoonzero Mar 17, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gangliao left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

helinwang left a comment

Choose a reason for hiding this comment

Yancey1989 left a comment

Choose a reason for hiding this comment

typhoonzero commented Mar 10, 2017 •

edited by wangkuiyi

Loading

gangliao commented Mar 10, 2017 •

edited

Loading

helinwang commented Mar 10, 2017 •

edited

Loading

gongweibao Mar 13, 2017 •

edited

Loading

typhoonzero Mar 16, 2017 •

edited

Loading

helinwang Mar 16, 2017 •

edited

Loading

helinwang Mar 16, 2017 •

edited

Loading

typhoonzero Mar 17, 2017 •

edited

Loading