This PR introduces single-node fault tolerance for Tune.
## Previous behavior:
- Actors will be restarted without checking if resources are available. This can lead to problems if we lose resources.
## New behavior:
- RUNNING trials will be resumed on another node on a best effort basis (meaning they will run if resources available).
- If the cluster is saturated, RUNNING trials on that failed node will become PENDING and queued.
- During recovery, TrialSchedulers and SearchAlgorithms should receive notification of this (via `trial_runner.stop_trial`) so that they don’t wait/block for a trial that isn’t running.
Remaining questions:
- Should `last_result` be consistent during restore?
Yes; but not for earlier trials (trials that are yet to be checkpointed).
- Waiting for some PRs to merge first (#3239)
Closes#2851.
* Suppress duplicate pre-emptive object pushes.
* Add test.
* Fix linting
* Remove timer and inline recent_pushes_ into local_objects_.
* Improve test.
* Fix
* Fix linting
* Enable retrying pull from same object manager. Randomize object manager.
* Speed up test
* Linting
* Add test.
* Minor
* Lengthen pull timeout and reissue pull every time a new object becomes available.
* Increase pull timeout in test.
* Wait for nodes to start in object manager test.
* Wait longer for nodes to start up in test.
* Small fixes.
* _submit -> _remote
* Change assert to warning.
IMPALA support for multiagent was broken since IMPALA has a requirement that batch sizes be of a certain length. However multi-agent envs can create variable-length batches.
Fix this by adding zero-padding as needed (similar to the RNN case).
* example
* add env
* test pg
* change to test
* add atexit test
* Update rllib-env.rst
* comment
* revert unnecessary file
* fix title when actor is idle
* Update python/ray/actor.py
Co-Authored-By: ericl <ekhliang@gmail.com>
This includes most of the TF code used for the OSDI experiment. Perf sanity check on p3.16xl instances: Overall scaling looks ok, with the multi-node results within 5% of OSDI final numbers. This seems reasonable given that hugepages are not enabled here, and the param server shards are placed randomly.
$ RAY_USE_XRAY=1 ./test_sgd.py --gpu --batch-size=64 --num-workers=N \
--devices-per-worker=M --strategy=<simple|ps> \
--warmup --object-store-memory=10000000000
Images per second total
gpus total | simple | ps
========================================
1 | 218
2 (1 worker) | 388
4 (1 worker) | 759
4 (2 workers) | 176 | 623
8 (1 worker) | 985
8 (2 workers) | 349 | 1031
16 (2 nodes, 2 workers) | 600 | 1661
16 (2 nodes, 4 workers) | 468 | 1712 <--- OSDI perf was 1817
Add new search algorithm (genetic) along with the base framework of the searcher (which performs some basic jobs such as logging, recording and organizing in our project).
Note that this is the initial commit. In the following days, we will add example, UT, and other refinements.
It's possible to configure PPO in a way that ends up discarding most of the samples (they are treated as "stragglers"). Add a warning when this happens, and raise an exception if the waste is particularly egregious.
A bunch of minor rllib fixes:
pull in latest baselines atari wrapper changes (and use deepmind wrapper by default)
move reward clipping to policy evaluator
add a2c variant of a3c
reduce vision network fc layer size to 256 units
switch to 84x84 images
doc tweaks
print timesteps in tune status
Rename AsyncSamplesOptimizer -> AsyncReplayOptimizer
Add AsyncSamplesOptimizer that implements the IMPALA architecture
integrate V-trace with a3c policy graph
audit V-trace integration
benchmark compare vs A3C and with V-trace on/off
PongNoFrameskip-v4 on IMPALA scaling from 16 to 128 workers, solving Pong in <10 min. For reference, solving this env takes ~40 minutes for Ape-X and several hours for A3C.
The dict merge prevents crashes when tune is trying to get resource requests for agents and you override a config subkey. The min iter time prevents iterations from getting too small, incurring high overhead. This is easy to run into on Ape-X since throughput can get very high.
This adds a simple DQN+PPO example for multi-agent. We don't do anything fancy here, just syncing weights between two separate trainers. This potentially is wasting some compute, but is very simple to set up.
It might be nice to share experience collection between the top-level trainers in the future.
Cleanup: TFPolicyGraph now automatically adds loss input entries for state_in_*, so that graph sub-classes don't need to worry about it.
Multi-GPU support:
Allow setting up model tower replicas with existing state input tensors
Truncate the per-device minibatch slices so that they are always a multiple of max_seq_len.
* Add profile table and store profiling information there.
* Code for dumping timeline.
* Improve color scheme.
* Push timeline events on driver only for raylet.
* Improvements to profiling and timeline visualization
* Some linting
* Small fix.
* Linting
* Propagate node IP address through profiling events.
* Fix test.
* object_id.hex() should return byte string in python 2.
* Include gcs.fbs in node_manager.fbs.
* Remove flatbuffer definition duplication.
* Decode to unicode in Python 3 and bytes in Python 2.
* Minor
* Submit profile events in a batch. Revert some CMake changes.
* Fix
* Workaround test failure.
* Fix linting
* Linting
* Don't return anything from chrome_tracing_dump when filename is provided.
* Remove some redundancy from profile table.
* Linting
* Move TODOs out of docstring.
* Minor
## What do these changes do?
**Vectorized envs**: Users can either implement `VectorEnv`, or alternatively set `num_envs=N` to auto-vectorize gym envs (this vectorizes just the action computation part).
```
# CartPole-v0 on single core with 64x64 MLP:
# vector_width=1:
Actions per second 2720.1284458322966
# vector_width=8:
Actions per second 13773.035334888269
# vector_width=64:
Actions per second 37903.20472563333
```
**Async envs**: The more general form of `VectorEnv` is `AsyncVectorEnv`, which allows agents to execute out of lockstep. We use this as an adapter to support `ServingEnv`. Since we can convert any other form of env to `AsyncVectorEnv`, utils.sampler has been rewritten to run against this interface.
**Policy serving**: This provides an env which is not stepped. Rather, the env executes in its own thread, querying the policy for actions via `self.get_action(obs)`, and reporting results via `self.log_returns(rewards)`. We also support logging of off-policy actions via `self.log_action(obs, action)`. This is a more convenient API for some use cases, and also provides parallelizable support for policy serving (for example, if you start a HTTP server in the env) and ingest of offline logs (if the env reads from serving logs).
Any of these types of envs can be passed to RLlib agents. RLlib handles conversions internally in CommonPolicyEvaluator, for example:
```
gym.Env => rllib.VectorEnv => rllib.AsyncVectorEnv
rllib.ServingEnv => rllib.AsyncVectorEnv
```
