Switches & Queues

Scope: Days AGO and Legacy Days. The global [switch] table configures every egress scheduler created from the topology.

[switch]
port_rate = 1000000000
capacity = 100
discipline = "FIFO"
drop = "TailDrop"

The strict legacy loader rejects inactive combinations. It permits the historical neutral weights = [1] token with FIFO, SP, or VirtualClock, but rejects other inactive weights.

Scheduling disciplines

All classed schedulers map class_id = flow_id % class_count.

For SP, priorities[class_id] gives the service priority; larger values run first, with FIFO order inside a priority. Missing legacy priorities default to [1]. Exact lowering uses the same default.

Legacy requires a non-empty weights array for WFQ, DRR, and WRR. Exact lowering also requires every weight to be positive. It converts DRR weights into 1500-byte-relative quanta. Metal and CUDA reject WFQ checkpoints whose rational numerator or denominator exceeds the device representation; the error reports the field and required bit width.

Legacy VirtualClock defaults missing vticks to [1.0]. Each tick is seconds per bit. Exact lowering returns:

unsupported scheduler `VirtualClock`; Days executor supports FIFO, SP, WFQ, DRR, and WRR

Drop and ECN policies

TailDrop drops a packet when admitting it would exceed packet capacity. Capacity 0 means unlimited for TailDrop.

Days AGO: the ECN ramp

Days AGO has one drop rule, tail drop at the queue's capacity, and one marking policy, the ECN ramp. It follows the Kmin/Kmax/Pmax shape of RED-style ECN on RoCE and DCQCN switches, over the instantaneous queue in bytes:

[switch]
port_rate = 100000000000
capacity = 3729
discipline = "FIFO"
drop = "TailDrop"
ecn_capacity_bytes = 33554432
ecn = { kmin_bytes = 400000, kmax_bytes = 1600000, pmax = 0.2 }

The decision is made once, when a packet arrives at a switch egress queue. Let d be the queue's bytes after admitting the packet (the packet in service is not counted):

1. If d exceeds ecn_capacity_bytes, the packet is dropped.
2. A packet that is not ECN-capable data is admitted unmarked, at any depth. ECN-capable data is plain, TCP and RoCE data; ACKs, NACKs, CNPs and PFC frames are never marked.
3. Below kmin_bytes the packet is admitted unmarked; at or above kmax_bytes it is marked.
4. In between, it is marked with probability pmax * (d - kmin) / (kmax - kmin).

The probability is decided by an exact integer test. With span = pmax_den * (kmax - kmin), the packet is marked when floor(u * span / 2^64) < pmax_num * (d - kmin), where u is a 64-bit draw. This equals u / 2^64 < P(d), so the probability is never rounded; the draw has a resolution of 2^-64. At d = kmin the probability is zero.

The draw is stateless: u = mix(queue_key ^ payload_id), where queue_key = mix(mix(mix(seed ^ ECN_RAMP) ^ node) ^ queue) and mix is SplitMix64's output function. Its inputs are the scenario seed, the switch egress port, and the packet's payload id (one per transmission attempt), so the draw does not depend on event order, worker count or backend, and no packet (an ACK included) can consume or reset another's draw. A retransmitted packet draws again.

pmax is an exact decimal in (0, 1], lowered to a fraction in lowest terms (0.2 is 1/5). The step is the special case kmin_bytes = kmax_bytes with pmax = 1: it marks every ECN-capable data packet at or above the threshold. Lowering refuses kmin_bytes of zero, kmin_bytes > kmax_bytes, kmax_bytes > ecn_capacity_bytes, a step with pmax below one, and a span beyond 64 bits.

The ramp runs on Scalar, CPU, Metal and CUDA with byte-identical results. On Metal and CUDA it is part of the mechanisms build of the round kernel, so an image without ECN runs the plain build, which admits by tail drop alone. LeanGuard's p10c_aqm_check recomputes every decision and its draw from the AQM certificate (see LeanGuard).

The SimAI fidelity lowers SimAI.conf's KMIN_MAP, KMAX_MAP and PMAX_MAP rows into ecn_by_rate (KB times 1,000), and BUFFER_SIZE into ecn_capacity_bytes; see Workloads. Days AGO decides at enqueue, where SimAI marks at dequeue; both decision points exist in switch hardware.

Legacy

ECN_THRESHOLD has a default fraction of 0.8. Legacy marks an ECN-capable packet and drops a non-ECN-capable packet at the threshold; it marks when post-admission depth is greater than floor(ecn_threshold * capacity). Physical capacity overflow drops packets.

Legacy RED and RED_ECN use packet occupancy, thresholds at 70% and 90% of capacity, EWMA weight 2^-9, and maximum probability 0.8. RED drops on a congestion action. RED_ECN marks an ECN-capable packet on that action and still drops at physical overflow.

ECN-capable transports

Legacy TCP emits Ect0 only when [flow.traffic.tcp] ecn = true. DCQCN data is ECN-capable in both engines. Days AGO TCP rejects ECN and models loss feedback only.

PFC interaction

PFC monitors per-priority byte occupancy while the queue keeps its own admission rule (packet capacity for TailDrop, ecn_capacity_bytes for the ECN ramp). Legacy schedulers expose shared queue state to their PFC ingress actors. Exact PFC runs on Scalar, CPU, Metal, and CUDA. See L2 and PFC.