Display exception when creating multiple SMARTS instances

[zbzhu99]

I encounter the "Display is not found" exception when trying to create multiple SMARTS instances:

To reproduce:

train_env = gym.make(
    "smarts.env:hiway-v0",
    scenarios=scenarios,
    agent_specs={AGENT_ID: agent_spec},
    headless=headless,
    visdom=False,
    timestep_sec=0.1,
    sumo_headless=True,
    seed=seed,
    # envision_record_data_replay_path="./data_replay",
)

test_env = gym.make(
    "smarts.env:hiway-v0",
    scenarios=scenarios,
    agent_specs={AGENT_ID: agent_spec},
    headless=headless,
    visdom=False,
    timestep_sec=0.1,
    sumo_headless=True,
    seed=seed,
    # envision_record_data_replay_path="./data_replay",
)

Development Environment:

OS: Ubuntu 16.04

Python: 3.7.4

SUMO: 1.7.0

SMARTS: 0.4.4

[Gamenot]

Hello @zbzhu99 , this is a known issue. SMARTS uses a few dependencies that have the requirement for a single instance in a process. In order to use another SMARTS instance in the same thread, the previous needs to be destroyed, or run on another process.

There is an example/multi_instance.py that shows one way to be able use more than one smarts environment using ray.

https://github.com/huawei-noah/SMARTS/blob/master/examples/multi_instance.py#L68-L162

[davidrusu]

@Gamenot We might be able to solve this problem using the suggestions in this comment here: panda3d/panda3d#996 (comment)

Namely, use the lower level components of panda3d instead of ShowBase

[Gamenot]

@davidrusu Yes, it seems like option 1 from that list of suggestions is the way to go eventually. We are already recommending option 3 via ray but stripping away the globals so that multiple instances can run in the same context would lower complexity for the users

[alek5k]

Hi there, I wrote the original github issue over at Panda3D. Just wondering if you'd be willing to share your implementation of the referenced option 1 from my github issue on Panda3D? I ended up going with option 3, but my training time is suffering :\

[davidrusu]

@alek5k We'd be happy to share it once we've got around to tackling this. Currently we are also relying on multiprocessing to avoid multiple instances of Showbase.

If you beat us to it, please come back and share your findings.

[alek5k]

Another relevant issue with Panda3D that you might run into if you are planning on running many instances in linux:

panda3d/panda3d#1064

[davidrusu]

@alek5k Yeah we did hit that one with one of our users as well

Maybe you can find something of use from the other configuration options we tuned here:

https://github.com/huawei-noah/SMARTS/blob/master/smarts/core/smarts.py#L62-L68

[Gamenot]

Since we are just using Showbase for the current option figured out here for us in particular is to use a Showbase subclass as a singleton a.k.a. option 2.

[Gamenot]

This particular problem has been addressed using option 2. The rest of the issue is covered by #743 .