Schedulers

Schedulers live in src/schedulers/ and model:

• queueing (single queue or multiple classes/queues),
• drop/mark behavior when congested,
• serialization delay at a given link rate,
• periodic reporting to CSV.

The topology builder creates one scheduler per directed edge in the topology graph (see Topology::connect_neighbours in src/topos/topo.rs).

Common structure

Most schedulers share these ideas:

• rate: f64: output bit rate (bits/s).
• busy_until: f64: time until which the server/link is “sending”.
• a queue (or set of queues) of Packet.
• a drop_strategy: Box<dyn PacketDrop> controlling enqueue/drop/ECN marking.

The typical control flow:

1. packet_received(packet) decides DropAction:
   • Drop: drop the packet
   • MarkEcn: attempt packet.mark_ce() (drop if not ECN-capable)
   • Enqueue: enqueue
2. If the server is idle (packet.time >= busy_until), start run(...).
3. run(...) pops a packet, updates queueing delay, computes a send timeout, schedules:
   • send(packet) after the timeout
   • another run(...) after the timeout (to keep draining the queue)

Scheduling disciplines

Configured via switch.discipline:

• FIFO: Port (src/schedulers/port.rs) — single FIFO queue
• DRR: DRRServer (src/schedulers/drr.rs) — deficit round-robin over classes
• WRR: WRRServer (src/schedulers/wrr.rs) — weighted round-robin over classes
• WFQ: WFQServer (src/schedulers/wfq.rs) — packet tagging + min-heap by finish time
• SP: SPServer (src/schedulers/sp.rs) — strict priority over classes
• VirtualClock: VirtualClockServer (src/schedulers/vc.rs) — virtual-clock tagging + min-heap

“Class” is defined by a closure flow_classes: Fn(flow_id) -> class_id. In the topology builder, this is commonly flow_id % num_classes where num_classes is weights.len() / priorities.len() / vticks.len().

Drop and ECN policies

Configured via switch.drop (src/schedulers/drop.rs):

• TailDrop: drop when capacity is exceeded
• RED: random early drop (RFC 2309)
• RED_ECN: RED, but marks ECN instead of dropping when possible
• ECN_THRESHOLD: mark CE when occupancy exceeds a configurable threshold, else enqueue

The switch config also supports:

• switch.ecn_threshold: a fraction in [0, 1] used by ECN_THRESHOLD (default: 0.8)

PFC integration (queue state)

When built with --features l2_pfc and run with link.mode="Pfc", schedulers expose a shared QueueState (src/schedulers/state.rs) that tracks current queued bytes/packets. This is used by the PFC ingress port to decide whether a packet can be forwarded downstream without causing drops.