FCFS load balancer (central queue) with Erlang-C validation, proper Wq measurement, and metrics plumbing

[GioeleB00]

What we changed

1. New FCFS implementation (central queue)

• Rewrote LB FCFS to a token-based FIFO:

  • LoadBalancerRuntime._free_edges: simpy.Store holding one token per free LB→server edge.
  • On dispatch, LB get()s one token (blocking if none), then forwards the request to that edge.
  • Servers call notify_server_free → lb.mark_free(edge_id) to return the token when they finish.
  • Zero “thundering herd”: only one waiting request wakes up per free server.
  • Added LoadBalancerRuntime.on_edge_added(edge_id) so EventInjection can re-enable edges cleanly (push a token when a server comes back).

2. Accurate FCFS waiting-time metric

• LB now records per-request waiting times for FCFS in _lb_waiting_time and exposes them via lb_waiting_times
• Arrival time at the LB is taken from the request’s last hop timestamp (history) at LOAD_BALANCER entry.
• Ensured link edges deliver on the next tick (yield env.timeout(0)) to preserve causal order of timestamps.

3. Results & analysis integration

• SimulationRunner.run() now passes the LB runtime into ResultsAnalyzer (lb=self._lb_runtime).

• ResultsAnalyzer exposes get_lb_waiting_times() and uses LB reference safely.

• MMc analyzer:

  • Detects FCFS and uses Erlang-C formulas for theory.
  • For observed Wq under FCFS, it now uses LB waiting times (central queue), not per-server CPU wait.
  • Compatibility checks cleaned up (LB must be FCFS for pooled queue; RANDOM/RR still treated as split model).

4. Event injection integration

• EventInjectionRuntime re-adds edges (server up) and calls lb_on_edge_added(edge_id) (wired from the runner) so the token is immediately available again.
• Removal of edges (server down) simply deletes from the LB’s OrderedDict; any stale tokens are discarded by the LB loop.

5. Unit tests

• New FCFS tests covering the token FIFO:

  • Immediate free edge ⇒ no wait recorded.
  • Wait until edge is added ⇒ Wq ≈ Δt.
  • Stale token discarded when edge removed, request waits for a valid edge.
  • FIFO order preserved across multiple edges & requests.

• Adjusted annotations (mypy) and lints (no blind except).