[sgd] add placement group support

[matthewdeng]

Why are these changes needed?

Placement groups can be used for many use-cases including improved load balancing, fault tolerance, and data colocation.

This change adds a default placement group behavior for a RaySGD worker group in which a single placement group is created with a bundle for each worker, and the SPREAD strategy is used for uniform distribution.

In the future, we may expand the external API to allow the user to define their own placement group strategy for more specific use-cases.

Note: In this implementation, if the worker is already in a placement group then it will be reused. This is done for the existing Tune-SGD integration in which the Tune Trainable will create a placement group to colocate workers for the Tune trials.

Related issue number

Closes #16682

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[amogkam]

looks good to me overall, just left a couple comments. Also don't forget to add your tests to the BUILD file

[amogkam]

LGTM, thanks! Can you ping again once CI passes

[HuangLED]

@matthewdeng @amogkam

Sorry for pinging in this 5-month old PR.

Got a question, in SGD v2, it is still the default behavior that trainer instances would use SPREAD strategy, correct? (saw something strange, would like to double-check first).

[matthewdeng]

Hey @HuangLED, apologies for not sharing this knowledge with you earlier. For Ray Train, we've actually decided to go with the alternative of defaulting to PACK. At the moment we expose a TRAIN_ENABLE_WORKER_SPREAD environment variable which can be set to 1 to use SPREAD. See #20091 for explanation in the description, and related code if you're interested.

[HuangLED]

apologies for not sharing this knowledge with you earlier. For Ray Train, we've actually decided to go with the alternative of defaulting to PACK. At the moment we expose a TRAIN_ENABLE_WORKER_SPREAD environment variable which can be set to 1 to use SPREAD. See #20091 for explanation in the description, and related code if you're interested.

Thank you! When saying "Ray Train", it refers to SGDV2, correct?

[matthewdeng]

@HuangLED yep, we've rebranded Ray SGD v2 as Ray Train!

Some useful references:

• Ray Train: Distributed Deep Learning
• Migrating from Ray SGD to Ray Train