GithubHelp home page GithubHelp logo

dataknower / hai-platform Goto Github PK

View Code? Open in Web Editor NEW

This project forked from hfailab/hai-platform

0.0 0.0 0.0 4.24 MB

一种任务级GPU算力分时调度的高性能深度学习训练平台

Home Page: https://hfailab.github.io/hai-platform/

License: GNU Lesser General Public License v3.0

Shell 5.41% Python 91.35% PLpgSQL 2.91% Dockerfile 0.33%

hai-platform's Introduction

HAI Platform

Header License

HAI 分时调度训练平台, 可通过 docker-compose 或 k8s 部署，提供功能：

  • 训练任务分时调度
  • 训练任务管理
  • jupyter 开发容器管理
  • studio用户界面
  • haienv 运行环境管理

外部依赖

  1. 一个集中存储，如 nfs, ceph, weka
    • 用于存放用户的运行代码
    • 代码运行输出的日志
    • 部署需要的 k8s conf
  2. k8s，确保你的运算主机都在 k8s 集群中
  3. 建议计算节点支持rdma，并安装 rdma-sriov device-plugin
    • 如果没有，在 配置项目 launcher.manager_envs 中配置 HAS_RDMA_HCA_RESOURCE: '0'

快速上手

  1. 构建

    构建 all-in-one hai-platform 镜像

    注：如需包含 haienv 202207 运行环境（包含cuda, torch），以同时作为训练任务镜像，需 export BUILD_TRAIN_IMAGE=1；如需自定义训练任务镜像，请参考 附录：初始化数据库 中 train_environment 的配置说明。

    # replace IMAGE_REPO with your own repo
$ IMAGE_REPO=registry.cn-hangzhou.aliyuncs.com/hfai/hai-platform bash one/release.sh
  build hai success:
    hai-platform image: registry.cn-hangzhou.aliyuncs.com/hfai/hai-platform:fa07f13
    hai-cli whl:
      /home/hai-platform/build/hai-1.0.0+fa07f13-py3-none-any.whl
      /home/hai-platform/build/haienv-1.4.1+fa07f13-py3-none-any.whl
      /home/hai-platform/build/haiworkspace-1.0.0+fa07f13-py3-none-any.whl

    安装 hai-cli 命令行

    pip3 install /home/hai-platform/build/hai-1.0.0+fa07f13-py3-none-any.whl
pip3 install /home/hai-platform/build/haienv-1.4.1+fa07f13-py3-none-any.whl
pip3 install /home/hai-platform/build/haiworkspace-1.0.0+fa07f13-py3-none-any.whl

    也可以使用预构建的镜像和命令行：

    # 仅包含 hai-platform
registry.cn-hangzhou.aliyuncs.com/hfai/hai-platform:latest
# 包含 hai-platform 和 haienv 202207 运行环境（包含cuda, torch）
registry.cn-hangzhou.aliyuncs.com/hfai/hai-platform:latest-202207

pip3 install hai --extra-index-url https://pypi.hfai.high-flyer.cn/simple --trusted-host pypi.hfai.high-flyer.cn -U

  2. 部署 hai-platform 到 k8s 集群

  • 获取使用帮助

    $ hai-up -h
Usage:
  hai-up.sh config/run/up/dryrun/down/upgrade [option]
  where:
    config:  print config script
    run/up:  run hai platform
    dryrun:  generate config template
    down:    tear down hai platform
    upgrade: self upgrade hai-cli/hai-up utility

    option:
      -h/--help:      show this help text
      -p/--provider:  k8s/docker-compose, default to k8s
      -c/--config:    show config scripts to setup environment variables,
                      if not specified, current shell environment will be used,
                      if not shell environment exists, default value in 'hai-up config' will be used

Setup guide
  step 1: ensure the following dependencies satisfied
    - a kubernetes cluster with loadbalancer and ingress supported
    - a shared filesystem mounted in current host and other compute nodes
    - for provider docker-compose: docker and docker-compose should be installed in current host
  step 2: "hai-up config > config.sh", modify environment variables in config.sh
  step 3: "hai-up run -c config.sh" to start the all-in-one hai-platform.

  • 按提示配置相关环境变量，如共享文件系统路径、节点分组信息、训练镜像、用户信息、挂载点等，确认无误后部署（如需自定义更多配置，可运行 hai-up dryrun 获取配置模板并自行修改，手动部署）

    hai-up config > config.sh
# customize config.sh

hai-up run -c config.sh

  • 使用 hai-cli 初始化和提交任务

    HAI_SERVER_ADDR=`kubectl -n hai-platform get svc hai-platform-svc -o jsonpath='{.status.loadBalancer.ingress[0].ip}'`

# TOKEN 为 USER_INFO 设置的token
hai-cli init ${TOKEN} --url http://${HAI_SERVER_ADDR}

# python文件默认需放在用户工作目录 ${SHARED_FS_ROOT}/hai-platform/workspace/{user.user_name}
# 如置于其他路径，需要在pg数据库storage表中添加相应挂载项
hai-cli python ${SHARED_FS_ROOT}/hai-platform/workspace/$(whoami)/test.py -- -n 1

  • 如需停用hai-platform，运行 hai-up down

附录：配置说明

hai-up run/dryrun 会创建 init.sql, override.toml 配置文件，以及部署到 k8s/docker-compose 的yaml文件，如有自定义配置需求，可自行修改。部分配置解释如下。

网络端口

默认打开如下端口：

  • 80: server
  • 5432: postgresql
  • 6379: redis
  • 8080: studio
  1. hai-platform 将启动一个 webservice，监听在 80 端口
  2. 内建的 pgsql 和 redis 需要对外访问，让 manager 能够访问到
    • 用户如果用自己搭建的外部 pgsql 和 redis，那么后两个端口可以不设置
  3. studio 为平台提供管理 UI 页面

k8s 相关配置

平台需要在集群创建 deployment,statefulset,pod,configmap 等资源，需要通过 rbac 为平台账号授权，或者直接赋予admin权限，对应的 kubeconfig 会被挂载到hai-platform的默认路径/root/.kube/config

节点分组

目前支持两类分组，TRAINING_GROUP, JUPYTER_GROUP，分别表示训练节点和jupyter节点。 需要指定分组名，以及分组内节点列表，如：

export MARS_PREFIX="hai-platform-one"
export TRAINING_GROUP="training"
export JUPYTER_GROUP="jupyter_cpu"
export TRAINING_NODES="cn-hangzhou.172.23.183.227"
export JUPYTER_NODES="cn-hangzhou.172.23.183.226"

设置分组信息后，启动脚本会自动配置 k8s node 的 label 为 ${MARS_PREFIX}_mars_group=${TRAINING_GROUP} 或 ${MARS_PREFIX}_mars_group=${JUPYTER_GROUP}; 并给 hai-platform 的 schduler 设置调度分组，同时设置数据库中的quota

挂载点

默认会挂载如下路径

  - '${HAI_PLATFORM_PATH}/kubeconfig:/root/.kube:ro'                        # k8s config，必须挂载
  - '${HAI_PLATFORM_PATH}/log:/high-flyer/log'                              # platform 日志
  - '${DB_PATH}:/var/lib/postgresql/12/main'                                # 持久化 pgsql
  - '${HAI_PLATFORM_PATH}/redis:/var/lib/redis'                             # 持久化 redis
  - '${HAI_PLATFORM_PATH}/workspace/log:${HAI_PLATFORM_PATH}/workspace/log' # 保存任务日志
  - '${HAI_PLATFORM_PATH}/log/postgresql:/var/log/postgresql'               # postgres 日志
  - '${HAI_PLATFORM_PATH}/log/redis:/var/log/redis'                         # redis 日志
  - '${HAI_PLATFORM_PATH}/init.sql:/tmp/init.sql'                           # 初始化的数据库语句，在这里添加相关的帐号、依赖等等配置
  - '${HAI_PLATFORM_PATH}/override.toml:/etc/hai_one_config/override.toml'  # override配置
  1. 挂载 k8s 配置文件
  2. 若需要，可以挂载 pgsql 的目录作为持久化目录，注意，容器内配置了 /var/lib/postgresql/12/main 作为数据目录
  3. 用户的任务日志目录，需要在共享文件系统上，该目录信息会被配置到 override.toml 的 experiment.log.dist 中

数据库配置

内建数据库

镜像内建了 pgsql 和 redis，并且通过 override.toml 配置相关信息。

  [database.postgres.primary]
  host = '${HAI_SERVER_ADDR}'
  port = 5432
  user = '${POSTGRES_USER}'
  password = '${POSTGRES_PASSWORD}'
  [database.postgres.secondary]
  host = '${HAI_SERVER_ADDR}'
  port = 5432
  user = '${POSTGRES_USER}'
  password = '${POSTGRES_PASSWORD}'
  [database.redis]
  host = '${HAI_SERVER_ADDR}'
  port = 6379
  db = 0
  password = '${REDIS_PASSWORD}'

初始化数据库

hai-platfrom 的用户信息，以及集群挂载信息都在 postgres 数据库中，可以通过数据库来控制用户的 quota、storage、权限等等， 初始化时会执行一次 init.sql, 初始化后如需修改数据库信息，请直接到数据中更改。 init.sql 模板示例如下：

INSERT INTO "storage" ("host_path", "mount_path", "owners", "conditions", "mount_type", "read_only", "action", "active")
VALUES
      -- marsv2 的脚本, 运行的依赖，不要改
      ('marsv2-scripts-{task.id}:suspend_helper.py', '/marsv2/scripts/suspend_helper.py', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:task_log_helper.py', '/marsv2/scripts/task_log_helper.py', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:validate_image.sh', '/marsv2/scripts/validate_image.sh', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:waiting_for_master.sh', '/marsv2/scripts/waiting_for_master.sh', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:waiting_pods_done.py', '/marsv2/scripts/waiting_pods_done.py', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:start_jupyter_with_ext.sh', '/marsv2/scripts/start_jupyter_with_ext.sh', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:hf_login_handler.py', '/marsv2/scripts/hf_login_handler.py', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-scripts-{task.id}:hf_kernel_spec_manager.py', '/marsv2/scripts/hf_kernel_spec_manager.py', '{public}', '{}'::varchar[], 'configmap', true, 'add', true),
      ('marsv2-entrypoints-{task.id}', '/marsv2/entrypoints', '{public}', '{}'::varchar[], 'configmap', true, 'add', true)
ON CONFLICT DO NOTHING;

INSERT INTO "storage" ("host_path", "mount_path", "owners", "conditions", "mount_type", "read_only", "action", "active")
VALUES
      -- 用户工作目录
      ('${HAI_PLATFORM_PATH}/workspace/{user.user_name}', '${HAI_PLATFORM_PATH}/workspace/{user.user_name}', '{public}', '{}'::varchar[], 'Directory', false, 'add', true)
ON CONFLICT DO NOTHING;

-- 和上面创建的分组对应起来 training
INSERT INTO "quota" ("user_name", "resource", "quota")
VALUES
      -- 设置的分组和权限，public 为所有用户，权限一共有 7 级
      -- LOW, BELOW_NORMAL, NORMAL, ABOVE_NORMAL, HIGH, VERY_HIGH, EXTREME_HIGH
      ('public', 'node-${TRAINING_GROUP}-HIGH', 10),
      -- 设置可用的训练节点数量
      ('public', 'node', 1),
      -- 系统内建镜像，和下面的 train_environment 对应起来
      ('public', 'train_environment:hai_base', 1),
      -- 任务容器的capabilities
      ('public', 'cap:IPC_LOCK', 1),
      -- jupyter任务的分组
      ('public', 'jupyter:${JUPYTER_GROUP}', 100012810)
ON CONFLICT DO NOTHING;

INSERT INTO "user" ("user_id", "user_name", "token", "role", "active", "shared_group")
VALUES
      -- 用户
      ('10020', 'testuser', 'xxxxxxxx', 'internal', true, 'hfai')
ON CONFLICT DO NOTHING;

INSERT INTO "train_environment" ("env_name", "image", "schema_template", "config")
VALUES
      -- 训练镜像，如自定义，需要满足 validate_image.sh 中镜像的条件
      ('hai_base', '${TRAIN_IMAGE}', '', '{"environments": {}, "python": "/usr/bin/python3.8"}')
ON CONFLICT DO NOTHING;

INSERT INTO "host" ("node", "gpu_num", "type", "use", "origin_group", "room")
VALUES
      -- 机器信息
      ('cn-hangzhou.172.23.183.226', 4, 'gpu', 'training', '${TRAINING_GROUP}', 'A'),
      ('cn-hangzhou.172.23.183.227', 4, 'gpu', 'training', '${TRAINING_GROUP}', 'A')
ON CONFLICT DO NOTHING;

平台配置

平台配置为 toml 格式，镜像中已经默认做了大量配置，用户只需要做基本的个性化配置即可，通过override.toml来覆盖镜像默认配置。 override.toml 示例：

  # 任务日志的位置
  [[experiment.log.dist]]
  role = 'internal'
  dir = '${HAI_PLATFORM_PATH}/workspace/log/{user_name}'
  # 数据库配置
  [database.postgres.primary]
  host = '${HAI_SERVER_ADDR}'
  port = 5432
  user = '${POSTGRES_USER}'
  password = '${POSTGRES_PASSWORD}'
  [database.postgres.secondary]
  host = '${HAI_SERVER_ADDR}'
  port = 5432
  user = '${POSTGRES_USER}'
  password = '${POSTGRES_PASSWORD}'
  [database.redis]
  host = '${HAI_SERVER_ADDR}'
  port = 6379
  db = 0
  password = '${REDIS_PASSWORD}'
  [scheduler]
  default_group = '${TRAINING_GROUP}'  # 提交任务时候的调度默认分组
  [launcher]
  api_server = '${HAI_SERVER_ADDR}' # 设定 server 地址
  task_namespace = '${TASK_NAMESPACE}'     # 任务运行的namespace
  manager_nodes = [${MANAGER_NODES_FORMATTED}] # manager 所在节点, 格式为['node1','node2']
  manager_image = '${BASE_IMAGE}'  # manager 使用的镜像，和 hai image 一致即可
  # 下面配置 manager 使用的 kube config
  [launcher.manager_envs]
  KUBECONFIG = '/root/.kube/config'
  HAS_RDMA_HCA_RESOURCE = '${HAS_RDMA_HCA_RESOURCE}'
  [launcher.manager_mounts]
  kubeconfig = '${HAI_PLATFORM_PATH}/kubeconfig:/root/.kube'

  [jupyter]
  shared_node_group_prefix='${JUPYTER_GROUP}'
  [jupyter.builtin_services.jupyter.environ.internal]
  JUPYTER_DIR = '${HAI_PLATFORM_PATH}/workspace/{user_name}/jupyter'
  [jupyter.builtin_services.jupyter.environ.external]
  JUPYTER_DIR = '${HAI_PLATFORM_PATH}/workspace/{user_name}/jupyter'
  [jupyter.ingress_host]
  internal = '${INGRESS_HOST}'
  external = '${INGRESS_HOST}'
  studio = '${INGRESS_HOST}'

SSH 配置

开发容器及 hai-cli 都提供了 SSH 相关的功能，但出于安全考虑目前任务容器中默认不启动 SSH 服务，也未为平台用户自动生成用于 SSH 连接的密钥。

如果想要使用相关功能，可以按照如下步骤配置：

  1. 编写 SSH 服务的初始化脚本，用于在任务容器的初始化阶段完成配置 SSH Key、启动 SSH 服务等工作；
  2. 修改数据库的 storage 表，将此脚本挂载至任务容器的 /usr/local/sbin/hf-scripts/post_system_init/ 目录下， 以使其在任务容器的初始化阶段运行。关于此目录的更多信息，可以参考 /marsv2/entrypoints 中的任务初始化脚本。

hai-platform's People

Contributors

vachelhu avatar mbalatsko avatar xyh0202 avatar mowentian avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.