Best-effort default means voice competes with backups. QoS is the pipeline: classify → mark → queue → drop smartly → shape/police, following DiffServ (IntServ/RSVP never scaled).

Marking

• L2 CoS/802.1p (0–7) on the frame, L3 DSCP (6 bits) in the IP header: voice EF(46), video AF41, data BE(0).
• Trust boundary: trust the phone, remark everything from a PC. Get this wrong and QoS is theatre.

Queuing

FIFO → WFQ → CBWFQ (bandwidth per class) → LLQ: a strict-priority queue for voice (policed so it can’t starve the rest). Congestion management = ordering; congestion avoidance = WRED dropping TCP early.

Police vs shape

Policing	Shaping
drops/remarks over rate, ingress, no buffering	buffers and delays to smooth, egress

class-map match-any VOICE
 match dscp ef
policy-map QOS
 class VOICE
  priority percent 10
 class class-default
  fair-queue
interface Gi0/0
 service-policy output QOS

QoS does nothing until there’s congestion — and then only if it’s end-to-end.