Distributed simulations

[diegoferigo]

This issue is somehow related to #12 about parallel simulations. The aim of #12 is exploiting multiple threads (or processes) running inside the same machine, here we focus more on exploiting multiple machines.

The obvious problem when executing distributed computations is synchronization between workers. The best solution that in the same time minimizes our effort (considering our current workforce) while maximizes the result is relying on ignition gazebo features:

• distributed_simulation.md
• levels.md

Currently these features are still not mature, but there are already working demos (distributed and distributed levels).

To be honest, in order to be able to scale a simulation horizontally (i.e. spawn transparently many parallel workers), probably the parallel simulations of #12 can be treated in the same way of distributed simulations running on a single machine. This would greatly simplify their maintenance. The tradeoff is that in this way #12 would be implemented as multi-process instead of multi-thread.

[diegoferigo]

As reported here:

The distribution of simulation work utilizes the concept of performers in order to set where physics simulation will occur. A performer is an additional annotation in an SDF file which marks each model that will be a performer. Performers are allocated to secondaries using a round-robin fashion. If there are more performers than secondaries, multiple performers will be allocated to each secondary.

A super cool feature to achieve would be running a simulation in a Kubernetes cluster, where each secondary is a pod and performers are instances of robot models.

[diegoferigo]

Note

Cleaning up old issues and pull requests I opened over the past years that are now outdated or stale. If this is still relevant, feel free to reopen it or create a new one with updated details. Thanks!