notes.md

Install ray_ir package locally

pip install -e ray_ir

Dependency IR

"Grammar"

Graph = Task(Str label) | Actor(Str label) | Map(Str label, Array args) -> Array | Broadcast(Str label, Array args) -> Array | InitActors(Str label, Array args) -> ActorArray | MapActors(Str label, ActorArray actors, Array args) -> ActorArray, Array | Repeat(Value v, Int n) -> Array

Value = Array | Object ActorArray = [ Actor ] Array = [ (Object, ...) ] Object = Task | Actor | python primitive

For(Graph *graphs, Int n) # The loop body needs to return a graph or list of graphs?

MapReduce example

reducers = InitActors("Reducer.init", num_reducers, args) dependencies = Task("generate_dependencies")

maps = Broadcast("map_step", dependencies) shuffles = Map("shuffle", maps) reducers = MapActors("Reducer.reduce", reducers, shuffles)

Example for using the scheduler interface:

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

# Submit a task normally.
f.remote(1)

# Submit the same task several times, but this time as a part of a group.
group_id = ray.ObjectID(ray.utils.random_string())
group = [f._remote(args=[1],
           group_id=group_id  # Which group this task is part of.
           ) for _ in range(10)]

# Submit another task that is dependent on the group.
f._remote(
  args=group,
  group_dependency=group_id  # Which group this task depends on.
  )

# Tell the scheduler that we no longer care about where the group was placed.
ray.free_groups([group_id])