does an instance of a Ray actor run in only one process?

Question

Can the methods of a Ray actor be called in parallel on a single remote instance? it seems i have to create a new instance of the actor, then call the method on that instance in order to achieve parallelization. See the code and out put below

@ray.remote
class ActorTest:
    def __init__(self) -> None:
        self.factor = 10
    
    def do_work(self, n):
        result = self.factor * n
        print(result)
        return result

if __name__ == '__main__':
    ray.init()
    actor = ActorTest.remote()
    results = [actor.do_work.remote(i) for i in range(10)]
    print(ray.get(results))

This code will print out the following - notice that the process ids are all the same:

(ActorTest pid=15176) 10
(ActorTest pid=15176) 20
(ActorTest pid=15176) 30
(ActorTest pid=15176) 40
(ActorTest pid=15176) 50
(ActorTest pid=15176) 60
(ActorTest pid=15176) 70
(ActorTest pid=15176) 80
(ActorTest pid=15176) 90

Now if i were to change the code to where i create a new remote instance each time i call my task method, it works as i would expect. each call is executed in a separate process as can be seen from the output:

@ray.remote
class ActorTest:
    def __init__(self) -> None:
        self.factor = 10
    
    def do_work(self, n):
        result = self.factor * n
        print(result)
        return result

if __name__ == '__main__':
    ray.init()
    results = []
    for i in range(10):
        actor = ActorTest.remote()
        results.append(actor.do_work.remote(i)) 
    print(ray.get(results))

Here you see the results are printed out with a different process id:

(ActorTest pid=31464) 70
(ActorTest pid=26272) 0
(ActorTest pid=35320) 20
(ActorTest pid=30868) 60
(ActorTest pid=10696) 10
(ActorTest pid=15688) 30
(ActorTest pid=27984) 90
(ActorTest pid=39852) 40
(ActorTest pid=10380) 80

Clarence Ng · Answer 1 · 2022-11-06T16:09:32.477

2

It is possible to call the methods of a single actor in parallel either via asyncio or threading: https://docs.ray.io/en/latest/ray-core/actors/async_api.html

As stated in the docs neither bypasses the Gil

edited Nov 06 '22 at 16:09

answered Nov 06 '22 at 16:09

Clarence Ng

31
2

ok..but not multi-processing though, correct? just confirming... – mike01010 Nov 06 '22 at 19:09
To do multi-processing it should create multiple actors, each actor instance is a process – Clarence Ng Nov 08 '22 at 09:45

does an instance of a Ray actor run in only one process?

1 Answers1