Ray

Ray

https://ray.io/

https://github.com/ray-project/ray

（1）机器学习生态基于python语言，但是python具有全局解释器锁缺点，限制了对单台机器的多核的利用

（2）同时查大规模模型的数据的出现，需要依赖集群来解决类似问题，引入了分布式机器学习的需求，

但是不需要引入更加高层的应用（spark）的基础上，ray基于python生态，单程的简单的分布式计算框架。

ray同时也包括了机器学习应用。

Ray provides a simple, universal API for building distributed applications.

Ray is packaged with the following libraries for accelerating machine learning workloads:

• Tune: Scalable Hyperparameter Tuning
• RLlib: Scalable Reinforcement Learning
• RaySGD: Distributed Training Wrappers
• Ray Serve: Scalable and Programmable Serving

https://docs.ray.io/en/latest/index.html

Ray provides a simple, universal API for building distributed applications.

Ray accomplishes this mission by:

1. Providing simple primitives for building and running distributed applications.

2. Enabling end users to parallelize single machine code, with little to zero code changes.

3. Including a large ecosystem of applications, libraries, and tools on top of the core Ray to enable complex applications.

https://www.ctolib.com/topics-138457.html

传统编程依赖于两个核心概念：函数和类。使用这些构建块就可以构建出无数的应用程序。

但是，当我们将应用程序迁移到分布式环境时，这些概念通常会发生变化。

一方面，OpenMPI、Python 多进程和 ZeroMQ 等工具提供了用于发送和接收消息的低级原语。这些工具非常强大，但它们提供了不同的抽象，因此要使用它们就必须从头开始重写单线程应用程序。

另一方面，我们也有一些特定领域的工具，例如用于模型训练的 TensorFlow、用于数据处理且支持 SQL 的 Spark，以及用于流式处理的 Flink。这些工具提供了更高级别的抽象，如神经网络、数据集和流。但是，因为它们与用于串行编程的抽象不同，所以要使用它们也必须从头开始重写应用程序。

用于分布式计算的工具

Ray 占据了一个独特的中间地带。它并没有引入新的概念，而是采用了函数和类的概念，并将它们转换为分布式的任务和 actor。Ray 可以在不做出重大修改的情况下对串行应用程序进行并行化。

来源（论文）

https://arxiv.org/abs/1703.03924

Real-Time Machine Learning: The Missing Pieces

Machine learning applications are increasingly deployed not only to serve predictions using static models, but also as tightly-integrated components of feedback loops involving dynamic, real-time decision making. These applications pose a new set of requirements, none of which are difficult to achieve in isolation, but the combination of which creates a challenge for existing distributed execution frameworks: computation with millisecond latency at high throughput, adaptive construction of arbitrary task graphs, and execution of heterogeneous kernels over diverse sets of resources. We assert that a new distributed execution framework is needed for such ML applications and propose a candidate approach with a proof-of-concept architecture that achieves a 63x performance improvement over a state-of-the-art execution framework for a representative application.

 

架构

https://www.cnblogs.com/fanzhidongyzby/p/7901139.html

论文给出的架构图里并未画出Driver的概念，因此我在其基础上做了一些修改和扩充。

Ray的Driver节点和和Slave节点启动的组件几乎相同，不过却有以下区别：

1. Driver上的工作进程DriverProcess一般只有一个，即用户启动的PythonShell。Slave可以根据需要创建多个WorkerProcess。
2. Driver只能提交任务，却不能接收来自全局调度器分配的任务。Slave可以提交任务，也可以接收全局调度器分配的任务。
3. Driver可以主动绕过全局调度器给Slave发送Actor调用任务（此处设计是否合理尚不讨论）。Slave只能接收全局调度器分配的计算任务。

https://zhuanlan.zhihu.com/p/41875076

其中的原理是将代码序列化到 redis 上存储为 object (object 可以理解为高效的不可变对象和数据共享)，实现各种异步执行和数据交换，优先在本地节点完成任务，如果完不成再由global scheduler 调配到其它节点（更正补充）。

DEMO CODE

单机版本，分布式任务示例。

remote声明函数为一个任务。

remote调用会将任务分发到一个计算进程中，并执行。

import ray
ray.init()

@ray.remote
def f(x):
    return x * x

futures = [f.remote(i) for i in range(4)]
print(ray.get(futures))

聚类学习工作流改造

https://github.com/fanqingsong/machine_learning_workflow_on_ray

from csv import reader
from sklearn.cluster import KMeans
import joblib
import ray


ray.init()


# Load a CSV file
def load_csv(filename):
    file = open(filename, "rt")
    lines = reader(file)
    dataset = list(lines)
    return dataset

# Convert string column to float
def str_column_to_float(dataset, column):
    for row in dataset:
        row[column] = float(row[column].strip())

# Convert string column to integer
def str_column_to_int(dataset, column):
    class_values = [row[column] for row in dataset]
    unique = set(class_values)
    lookup = dict()
    for i, value in enumerate(unique):
        lookup[value] = i
    for row in dataset:
        row[column] = lookup[row[column]]
    return lookup

def getRawIrisData():
    # Load iris dataset
    filename = 'iris.csv'
    dataset = load_csv(filename)
    print('Loaded data file {0} with {1} rows and {2} columns'.format(filename, len(dataset), len(dataset[0])))
    print(dataset[0])
    # convert string columns to float
    for i in range(4):
        str_column_to_float(dataset, i)
    # convert class column to int
    lookup = str_column_to_int(dataset, 4)
    print(dataset[0])
    print(lookup)

    return dataset

@ray.remote
def getTrainData():
    dataset = getRawIrisData()
    trainData = [ [one[0], one[1], one[2], one[3]] for one in dataset ]

    return trainData

@ray.remote
def getNumClusters():
    return 3

@ray.remote
def train(numClusters, trainData):
    print("numClusters=%d" % numClusters)

    model = KMeans(n_clusters=numClusters)

    model.fit(trainData)

    # save model for prediction
    joblib.dump(model, 'model.kmeans')

    return trainData

@ray.remote
def predict(irisData):
    # test saved prediction
    model = joblib.load('model.kmeans')

    # cluster result
    labels = model.predict(irisData)

    print("cluster result")
    print(labels)


def machine_learning_workflow_pipeline():
    trainData = getTrainData.remote()
    numClusters = getNumClusters.remote()
    trainData = train.remote(numClusters, trainData)
    result = predict.remote(trainData)

    result = ray.get(result)
    print("result=", result)



if __name__ == "__main__":
    machine_learning_workflow_pipeline()

Ray 破冰学习

https://github.com/anyscale/academy/blob/master/ray-crash-course/00-Ray-Crash-Course-Overview.ipynb