Pin a Ray actor to a core

Ray Core
1

Is this even possible?

2

Hey @marsupialtail , would you mind elaborating a bit more on the context? Specifically, when you say pinning, what do you mean here?
E.g.

  • always runs an actor to a cpu by id?
  • always run an actor to one of cpus across programs?
  • always runs ActorA to the same core, and ActorB class to another?

I think there is no direct top level API in ray that pins actor to a core with some identifiers. e.g. ActorClass.remote(pin_cpu=<some_cpu_id>)

But there might be ways to make an actor to be scheduled on a worker(core) with placement groups or detached actors.

3

Yes I want to always run an actor on a particular CPU. In particular the actor has some state, so I want the state to stay in L2.
This is probably less important to the Python API but when using the C++ API I am thinking about this.

4

I’ve gotten this to work in Python before by simply running os.sched_setaffinity within the actor. Perhaps there’s an equivalent of that in CPP?

1 Like
5

Yes, you can totally do that on your own in c++ by using sched_setaffinity system call (sched_setaffinity(2) - Linux man page)