使用ray并行化模拟器python

问题描述

我是ray的新手,我正在尝试并行化我开发的模拟器。这是我的模拟器的一个例子,显然它更复杂。

import some_library
import sim_library_with_global_object

class Model(object):
    def __init__(self,init_vals):
        #initialize object using some of the global_object from sim_library.
        #the Model object have it's own variables not global

    def do_step(self,time):
        #calculate Model step using the global_object from sim_library
        #edit the Model variables with respect to the step


class ManyModel(object):
    def init(self):
        self.models=[]

    def add_model(self,init_vals):
        model = Model(init_vals)
        self.model.append(model)

    def step(self,time):
        for model in self.models:
            model.do_step(time)

    def get_data_step(self):
        data=[]
        for model in self.models:
            data.append(model.myvalues)
        return data



sim=ManyModel()
inits=[] #####list of init_vals
times=[] ####list of times to simulate
for init in intis:
    sim.add_model(init)

for time in times:
    sim.step(time)
    step_data=sim.get_data_step()

到目前为止,我已经尝试通过以下两种方式在@ray.remote类(1)和 Model类(2)上使用带有装饰器ManyModel的ray:

(1)

############################## (1) ###############
import some_library
import sim_library_with_global_object

@ray.remote
class Model(object):
    def __init__(self,time):
        #calculate Model step using the global_object from sim_library
        #edit the Model variables with respect to the step


class ManyModel(object):
    def init(self):
        self.models=[]


    def add_model(self,init_vals):
        model = Model.remote(init_vals)
        self.model.append(model)

    def step(self,time):
        futures=[]
        for model in self.models:
            futures.append(model.do_step.remote(time))
        return futures

    def get_data_step(self,futures):
        data=[]
        while len(futures)>0:
            ready,not_ready = ray.wait(ids)
            results=ray.get(ready)
            data.append(results)
        return data

ray.init()
sim=ManyModel()
inits=[] #####list of init_vals
times=[] ####list of times to simulate
for init in intis:
    sim.add_model(init)

for time in times:
    sim.step(time)
    step_data=sim.get_data_step()

(2)

########################## (2) #################

import some_library
import sim_library_with_global_object

class Model(object):
    def __init__(self,time):
        #calculate Model step using the global_object from sim_library
        #edit the Model variables with respect to the step

@ray.remote
class ManyModel(object):
    def init(self):
        self.models=[]
        self.data=[]


    def add_model(self,time):
        for model in self.models:
            model.do_step(time)

    def get_data_step(self):
        self.data=[]
        for model in self.models:
            self.data.append(model.myvalues)
        return self.data


ray.init()
sim=ManyModel.remote()
inits=[] #####list of init_vals
times=[] ####list of times to simulate
for init in intis:
    sim.add_model.remote(init)

for time in times:
    sim.step.remote(time)
    future=sim.get_data_step.remote()
    step_data=ray.get(future)

在两种方式下,使用ray库都没有任何好处。你能帮我吗?

方法(1)的更新 第一种方法的问题是我收到警告消息

2020-11-09 11:33:20,517 WARNING worker.py:1779 -- WARNING: 12 PYTHON workers have been started. This Could be a result of using a large number of actors,or it Could be a consequence of using nested tasks (see https://github.com/ray-project/ray/issues/3644) for some a discussion of workarounds.

对于10 x Model ,这是性能结果: 不使用射线: 10 x Model -> do_step 0.11 [s] 使用ray(1): 10 x Model -> do_step 0.22 [s]

此外,每次我使用方法(1)创建一个Actor时,它都会为导入的库创建所有global_objects的副本,并且使ram消耗变得疯狂。我需要用超过100k个Model 对象进行午餐模拟。

总的来说,我不知道在ray中创建很多演员是否是个好主意。

解决方法

放大一些核心元素

ray.init()
sim=ManyModel.remote()

for time in times:
    sim.step.remote(time)
    future=sim.get_data_step.remote()
    step_data=ray.get(future)

最重要的一点是,您仅创建一个Ray演员(在sim=ManyModel.remote()行中)。 Ray actor按顺序执行提交给他们的任务(默认情况下),因此创建一个actor不会为并行性创造任何机会。要与Ray actor保持并行,您需要创建并使用多个actor。

第二点是您正在for循环内调用ray.get。这意味着在通过for循环的每次迭代中,您都将提交一个任务,然后调用ray.get,等待它完成并检索结果。相反,您将要提交多个任务(可能在循环内),然后在循环外调用ray.get