Commit 5ca3a45e authored by Brewer, Wes's avatar Brewer, Wes
Browse files

fix(simulation): Correct adaptive routing logic and config 

The adaptive routing simulation was flawed because the network link loads
were not being reset at each timestep. This caused routing decisions to
be based on stale, cumulative data.

This commit fixes the issue by calling  at the
beginning of each simulation tick in .

Additionally, the  configuration is corrected
to be compatible with replay simulations, changing the policy to 'replay'
and removing the conflicting 'arrival' parameter.
parent a121e132

experiments/lassen.yaml

+1 −2
Original line number Diff line number Diff line 
system: lassen

replay:

  - /opt/data/lassen/Lassen-Supercomputer-Job-Dataset

policy: fcfs

policy: replay

backfill: firstfit

fastforward: 365d

time: 12h

arrival: poisson

simulate_network: true

raps/engine.py

+3 −0
Original line number Diff line number Diff line
@@ -548,6 +548,9 @@ class Engine: 
        is simulated.

        """



        if self.simulate_network and self.running:

            self.network_model.reset_link_loads()



        scheduled_nodes = []

        cpu_utils = []

        gpu_utils = []