paddle cloud 计划内容

[typhoonzero]

第一阶段： 可以演示，可以开放给部分用户使用(内测)

时间点：2017-05-31

• 只需要支持jupiter notebook on PaddlePaddle cloud when doing distributed training.
  • Jupiter notebook需要能收到callback，包含cost，能自己画图。（就跟单机版使用方式类似，都是callback里绘图）。
  • PaddlePaddle Server --yanxu （如果第一版只支持jupiter notebook on cloud，这个就不一定需要了）
    End user's train function talks to PaddlePaddle server, which invokes Docker to build images.
• Paddle Cloud Web页面 -- 先出原型-- wuyi, yanxu,gongweibao
  • 在网页上编写训练代码
  • 在云端打包并提交训练任务
  • training过程可视化（cost动画）
  • job运行状态监控
  • 查看训练日志
  • 查看个人配额
  • 支持inference应用（暂不考虑）
• RBAC--wuyi
  • 使用百度账号登录（注册）-->开通账号-->配置namespace等
• 存储GlusterFS -- weibao
  • 权限和配额
  • 训练数据的上传和分片
  • 性能考虑，确定演示版本部署方案。使用内部存储系统如BFS也会带来和kubernetes的适配成本
• pserver和trainer支持扩容的功能的开发
• Network Policy网络隔离调研
  • 不影响演示，但是需要有，否则不安全
• GPU资源 -- Done
• v2 分布式训练：强调扩容，体现出变化。
  • 训练任务扩容
    • 扩容之后体现效果展示？

• Implement master program. (helin)
  • master，trainer，pserver service discovery. (helin)
  • master trainer communication. (helin)
• Implement fault tolerant parameter server.
  • Do we need to rewrite parameter server? How much effort is it to add fault tolerant in C++? If the effort is bigger or equal to rewrite in golang, maybe we can rewrite in golang.
  • Do we need to support sparse parameter update in v1?
  • What kind of update rule does parameter server need to support in v1? maybe only simple "add" (no momentum based).
• Implement fault tolerant trainer.
  • able to scale up trainer.
  • it involves changes for python part and native part (c++ or golang). We need to define a clean C api for python to use.
• Setup etcd service (no detail specification in the design doc, will not reuse etcd from k8s, according to distributed training: should we re-use etcd from k8s? #1807).
  • How to control etcd access namespace?
• Upload custom dataset to cluster.
  • Do we need to support merge data files to a big custom file to speed up sequential read. This is performance in the first version?
  • How can the trainer read the dataset, and be backward / forward compatible, maybe we need a reader "driver" for each dataset?

风险：
- 长期考虑，高性能存储需要深度支持。
- Web页面的开发有工作量，人员少

---

Design docs:

• https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/cluster_train:
  • Fault tolerant master program
  • Fault tolerant parameter server
  • Fault tolerant trainer
  • Parameter server checkpointing / recover from checkpoint file.
  • Upload custom dataset to cluster / how can the reader use the uploaded dataset.
• Design doc: submit a distributed job #1770 (PR review in progress)
  • How to submit cluster training job from client.
  • PaddlePaddle Server

---

4/24/2017 meeting minutes:
scope information for first version:
pserver：
⁃ 只考虑TCP，不支持RDMA
⁃ 不考虑sparse
⁃ 支持trainer动态伸缩
⁃ 同步 SGD

trainer：
⁃ pserver client
⁃ fetch taskid，按task处理数据
⁃ 动态伸缩，demo强调扩容，体现出变化。

master：
⁃ 服务发现
⁃ 分配task

paddle server：
⁃ build docker image on Kubernetes
⁃ 启动paddle job

paddle client：
⁃ 提交集群任务（python代码, add an optional argument for paddle.train, which contains dist train configuration.）
⁃ 命令行 paddle upload/download

• 不允许用户在分布式训练里画图，只能打印log。
• Paddle会提供cost的动态图表。
• Parameter Server是否需要重写需要更多调研。
• PR问题达成了一致。
• 分工（请参考issue评论，4/24/2017的plan）

[jacquesqiao]

第一阶段，容错应该是个重要的点

“容错” 改为 “扩容并体现性能变化”

[helinwang]

4/24/2017 plan:

• prototype master, trainer, pserver service discovery through etcd in golang.
• look into if need to rewrite parameter server and trainer in golang.

[jacquesqiao]

我来调研一下parameter server 的情况吧

[typhoonzero]

• paddle cloud和k8s用户认证统一(使用openid)
• kubernetes RBAC用户权限
• 使用python + django + bootstrap实现web页面登录和启动jupyter notebook的原型
• 绘制训练cost的动态图像

[Yancey1989]

• 在python代码中提交集群任务
• Paddle server打包docker image，启动训练任务。

[gongweibao]

• Setup etcd service (no detail specification in the design doc, will not reuse etcd from k8s, according to distributed training: should we re-use etcd from k8s? #1807).

  • How to control etcd access namespace?

• Upload custom dataset to cluster.

  • 权限和配额
  • 训练数据的上传、下载
  • Do we need to support merge data files to a big custom file to speed up sequential read. This is performance in the first version?
  • How can the trainer read the dataset, and be backward / forward compatible, maybe we need a reader "driver" for each dataset?

• Implement fault tolerant trainer.

  • able to scale up trainer.
  • it involves changes for python part and native part (c++ or golang). We need to define a clean C api for python to use.

不确定：

• paddlecloud web开发

[dzhwinter]

4/24/2017 plan:
1、look into the parameter server detail, if we need to add fault tolerant in C++, auto scaling of parameter server process.
2、add parameter server checkpointing / recover from checkpoint file

[helinwang]

4/26/2017 plan

1. master's detailed design doc.
2. prototype master, train, pserver communication with etcd.
3. file server with cephfs in golang.

[typhoonzero]

Closing this issue, we can track the status in "Project": https://github.com/PaddlePaddle/Paddle/projects/18