Adding basic support for a user-interpretable resource label

[atumanov]

This PR provides the ability to configure an arbitrary resource per local scheduler and lets tasks request it. It natively supports infinite capacity out of the box.

This is an experimental first pass at addressing #695. There will be API changes down the road.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1422/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test FAILed.

[AmplabJenkins]

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1423/
Test FAILed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1424/
Test PASSed.

[robertnishihara]

Want to add some tests showing how to use this?

[atumanov]

yeah, I'll add some unit tests to exercise this. I've been testing it like this:

#!/usr/bin/env python
import ray
import time

@ray.remote(num_uirs=1)
def f():
  time.sleep(10)
  return 1

@ray.remote(num_uirs=1)
def g():
  return 2

ray.init(num_uirs=1)

oid = f.remote(); oids = [g.remote() for _ in range(10)]
t1 = time.time(); results = ray.get(oids); t2 = time.time()
print t2-t1, results

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1568/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1571/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1572/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1573/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1574/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1575/
Test PASSed.

[pcmoritz]

instead of unstable, let's say "support for this will be removed" or "experimental"; unstable has more a connotation of not working reliably

[mitar]

"will be removed" sounds like that it has been deprecated. I hope it will not just be removed, but maybe matured into something else.

Maybe just saying "experimental feature subject to changes in the future". Kubernetes has this for GPU support: https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/

[pcmoritz]

yeah, you are right, sounds good!

[mitar]

BTW, your GitHub tagline is "An experimental distributed execution engine". So this is then experimental experimental feature? ;-)

Maybe you should find a better tagline. Like "an awesome distributed execution engine". :-)

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/1577/
Test PASSed.

[mitar]

Thanks!

[robertnishihara]

This can be used as follows (for example).

Start three machines like this.

ray start --head --redis-port=6379

ray start --redis-address 172.31.10.143:6379 --num-custom-resource=10

ray start --redis-address 172.31.10.143:6379

Define a remote function that uses some of the "custom resource"

import ray

ray.init(redis_address="172.31.10.143:6379")

@ray.remote(num_custom_resource=1)
def f():
    import time
    time.sleep(0.01)
    return ray.services.get_node_ip_address()

print(set(ray.get([f.remote() for _ in range(1000)])))

The print statement should show that it only is scheduled on the second machine. Note that to start a machine with infinite "custom resource", you can use --num-custom-resource=-1.