[Ray component: Core] Respect the placement group when num_gpus=0, num_cpus=0

[JiahaoYao]

What happened + What you expected to happen

Goal: Ideally, we want the ray remote actors or tasks to respect the placement group bundle index even when no resources is allocated by ray.

Observation: Currently, the ray actor does not respect the ray placement group when no resources are specificed; and in order to respect the placement group, one might have to manually use num_cpus=1e-4.

@rkooo567

Versions / Dependencies

Ray 2.0

Reproduction script

# Import placement group APIs.
from ray.util.placement_group import (
    placement_group,
    placement_group_table,
    remove_placement_group
)
from ray._private.services import get_node_ip_address
from ray.util.placement_group import get_current_placement_group, remove_placement_group



# Initialize Ray.
import ray
from icecream import ic
ray.init('auto')



bundle1 = {"GPU": 1, "CPU": 1}
bundle2 = {"GPU": 1, "CPU": 1}

pg = placement_group([bundle1, bundle2], strategy="STRICT_SPREAD")



ray.get(pg.ready())


ic(ray._private.state.state._available_resources_per_node())


@ray.remote(num_gpus=0, num_cpus=0)
def f():

    import time 
    print(get_node_ip_address())
    time.sleep(10)
    return True

ray.get([f.options(placement_group=pg, placement_group_bundle_index=1).remote() for i in range(10)])
ray.get([f.options(placement_group=pg, placement_group_bundle_index=0).remote() for i in range(10)])

Result:

with num_cpus=0 , both outputs are the same ip

---

@ray.remote(num_gpus=0, num_cpus=1e-3)
def f():

    import time 
    print(get_node_ip_address())
    time.sleep(10)
    return True

ray.get([f.options(placement_group=pg, placement_group_bundle_index=1).remote() for i in range(10)])
ray.get([f.options(placement_group=pg, placement_group_bundle_index=0).remote() for i in range(10)])

Result:

with 1e-3 , the two are different

Issue Severity

No response

[hora-anyscale]

Per Triage Sync: @rkooo567 can you confirm if this was fixed?

[rkooo567]

This is not fixed.

[rkooo567]

Should be fixed by 2.3

[cadedaniel]

This will fix an issue we observed when measuring dataset preprocessing https://anyscaleteam.slack.com/archives/C041D0LT8ET/p1670628383549839

[jeremi-eta]

Hi, is this issue fixed? And if so, in which version of Ray? Thanks