展开查看详情
1 .Ray: 机器学习应用的计算底盘
张喆 / Anyscale
2021 / 4月17日
2 .Agenda
我们为什么要开发Ray?
Ray的架构和设计
Ray的应用场景
3 .我们为什么要开发Ray?
分布式计算必然走向主流
Ray：让分布式计算变得简单
4 .分布式系统的历史
Big Data
(2000s)
Web
(1990s)
HPC
(1980s)
ARPA Network
(1970s)
5 .分布式计算仍然是“小众”
对大部分开发者仍很陌生
学习曲线还是很陡峭
现在是分布式计算走向主流的时间节点
6 .分布式计算背后的关键趋势
深度学习的快速崛起
摩尔定律的终结
应用逻辑和高算力(AI/ML)的融合
7 . s
o nth
18m
v ery
xe
35
CPU
(https://openai.com/blog/ai-and-compute//)
8 .越来越多的应用需要数据和智能的支持
深度学习
科学计算
+
运筹优化
? 微服务
大数据
10 . serverless
Limitations
● cloud specific
● stateless only
开发成本
● no GPUs/TPUs
● runtime limit
stitch together
existing frameworks
Hard to
● develop
● deploy
● manage
clean slate
Expensive
to develop
● time
● people
开发通用性/灵活性
11 .GitHub Stars
Spark
Kafka
Ray
MLFlow
Dask
12 .Agenda
我们为什么要开发Ray?
Ray的架构和设计
Ray的应用场景
13 .追求像“普通”Python一样简单的API
Function / 函数 Task / 任务
Class / 类 Actor
Object / 对象 Object / 对象
15 . Ray “Hello World”
@ray.remote(num_cpus=1) Worker
class Worker: Actor
def work(self):
return "done" Supervisor
Driver
Actor Worker
@ray.remote(num_cpus=1) Actor
class Supervisor:
# Supervisor can do things like listing files in a dir
Worker
def __init__(self):
Actor
self.workers = [Worker.remote() for _ in range(3)]
def work(self):
return ray.get([w.work.remote() for w in
self.workers])
ray.init()
sup = Supervisor.remote()
print(ray.get(sup.work.remote()))
# outputs ['done', 'done', 'done']
16 . Ray “Hello World”
Worker
Actor
@ray.remote(num_cpus=1)
class Worker: Supervisor
def work(self): Driver
Actor Worker
return "done"
Actor
@ray.remote(num_cpus=1)
class Supervisor:
Actors are extremely powerful. They allow you to
# Supervisor can do things like listing files in a dir
Worker
Actor
take a Python class and instantiate it as a
def __init__(self):
self.workers = [Worker.remote() for _ in range(3)]
microservice which can be queried from other actors
def work(self):
return ray.get([w.work.remote() for w in self.workers])
and tasks and even other applications.
ray.init()
sup = Supervisor.remote()
print(ray.get(sup.work.remote()))
# outputs ['done', 'done', 'done']
19 . Ray的生态系统
Native Libraries Third Party Libraries
Ray becoming the go-to framework for
scaling libraries
20 .Agenda
我们为什么要开发Ray?
Ray的架构和设计
Ray的应用场景 -- 机器学习的计算底盘
21 .应用案例：Uber的机器学习系统
深度学习 经典 ML 无代码AutoML
容器调度系统
23 .基于Ray的新架构
在这里，Ray尝试找到“正确的抽象层”
Ray
Peloton (Resource Scheduler)
24 .Uber: By leveraging Ray, we can combine the preprocessing, distributed training,
and hyperparameter search all within a single job running a single training script.
25 .Uber: We believe that Ray will continue to play an increasingly important role in bringing
much needed common infrastructure and standardization to the production
machine learning ecosystem, both within Uber and the industry at large.
26 . 谢谢！欢迎加入Ray的社区！
• 安装 ，上手试一试！pip install ray
• 注册 Ray Summit!
https://www.anyscale.com/ray-summit-2021
• 浏览文档: docs.ray.io
• 关注项目进展：https://github.com/ray-project/ray
热门城市
