Optimization: server_on_change_hosts_pstate and server_on_turn_onoff_hosts early drops machines which already are at the target pstate.

[marcodamico]

Ported #75 to main branch.

[bleuse]

Hi @marcodamico, thanks for rebasing your PR to the main branch.

I am not sure this only an optimization: by not forwarding the seemingly useless pstate change, it seems to me we are changing what is simulated, and the visible behavior.

Could you detail why you think this change should be made here rather than checking in the EDC if an actual pstate change is required?

[marcodamico]

The idea is to skip the add_pstate_change functions for nodes that don't need a recalculation. This improved my simulations by a factor of 2.5x on 10000 nodes. I am not sure what you mean by EDC

[Mommessc]

EDC is the External Decision Component, usually it is the scheduler process.

[marcodamico]

Even if we filter the pstates in the EDC, batsim will keep calling add_pstate_change on all the machines.

[bleuse]

Looking at the code of add_pstate_change, this should not be the culprit for your performance issues as its cost is fairly constant on the number of machines.
Do you have any data related to this part, e.g., a flame graph?

Furthermore, I have the intuition that the speedup this PR achieves is coming from the fact you are requesting less operations (e.g., set_pstate) and triggering less logging from Simgrid.

As far as I understand what you are trying to do, your EDC/scheduler is requesting too many changes.
You should filter out irrelevant machines in the sent message.

Are you able to share the relevant parts of you EDC/scheduler code to help us understand what you are trying to achieve?