QoS service models
| Model | Idea | Operational note |
|---|---|---|
| Best effort | All traffic receives ordinary forwarding without differentiated guarantees. | Simple and sufficient when links never congest. |
| Integrated Services | Applications reserve resources per flow, commonly associated with RSVP. | Fine-grained but can create state and scaling complexity. |
| Differentiated Services | Traffic is marked into classes and receives a per-hop behaviour. | Widely used in enterprise and provider QoS designs. |
A QoS design begins with business requirements: which applications matter, what delay/loss/jitter they tolerate, where congestion occurs and how much capacity each class should receive.
The QoS toolkit
| Mechanism | Purpose |
|---|---|
| Classification | Identify traffic by fields, application recognition, interface, user or policy. |
| Marking | Set a reusable class indicator such as DSCP or 802.1p/PCP. |
| Queuing/scheduling | Decide which queued packets transmit first and how bandwidth is shared. |
| Congestion avoidance | Drop/mark selected packets before queues completely fill, often to influence TCP behaviour. |
| Policing | Enforce a rate by forwarding, remarking or dropping excess traffic. |
| Shaping | Buffer excess traffic and release it at a configured average rate. |
| Link-efficiency mechanisms | Reduce serialization impact or overhead where relevant. |
DSCP and trust boundaries
The IPv4 DS field and IPv6 Traffic Class carry the 6-bit DSCP value used for differentiated service. Common conventions include Default Forwarding (DSCP 0), Expedited Forwarding for tightly controlled low-loss/low-delay service, and Assured Forwarding classes.
A trust boundary is where the network decides whether to accept an endpoint's markings. Trust too early and users can mark bulk traffic as priority; trust too late and important application identity may be lost. Classify and mark as close to the source as operationally safe.
Delay, jitter, loss and queuing
- Serialization delay: time to place bits on the link; larger packets and slower links increase it.
- Propagation delay: time for the signal to travel through the medium.
- Processing delay: device inspection and forwarding work.
- Queuing delay: waiting behind other packets; highly variable during congestion.
- Jitter: variation in packet delay, important for real-time media.
- Loss: can result from full queues, policing, physical errors or policy.
QoS has its greatest effect at bottlenecks. Applying a sophisticated policy on an uncongested 10-Gbps link does not repair a congested 100-Mbps WAN edge unless the policy is placed where the queue actually forms.
Policing versus shaping
| Area | Policing | Shaping |
|---|---|---|
| Excess traffic | Usually dropped or remarked immediately. | Buffered and transmitted later. |
| Traffic profile | Can be bursty after enforcement. | Smooths the output rate. |
| Typical use | Enforce a contract or protect resources. | Match a downstream rate and reduce provider policer drops. |
| Added delay | Little buffering delay from the policer itself. | Can add delay because packets wait in the shaping queue. |
Policy example and troubleshooting
class-map match-any REALTIME match dscp ef ! policy-map WAN-EDGE class REALTIME priority percent 10 class CRITICAL-DATA bandwidth percent 30 class class-default fair-queue ! interface GigabitEthernet0/1 service-policy output WAN-EDGE
- Measure the interface and subinterface where congestion actually occurs.
- Verify classification and marking counters.
- Check whether markings survive switches, tunnels, firewalls and provider boundaries.
- Inspect queue depth, drops, policer exceed actions and shaping rate.
- Compare configured bandwidth with real circuit and provider committed rates.
- Use application metrics for delay, jitter, loss and retransmission—not only ping.
- Reassess class allocation if priority consistently consumes its limit.
Interview focus
Use these points to practise concise, operations-focused answers. Explain the concept first, then give a verification or troubleshooting example.
- Can QoS create more bandwidth? No. It manages how available resources are used during contention.
- What is the difference between policing and shaping? Policing drops or remarks excess traffic; shaping buffers it for later transmission.
- Where should QoS marking be trusted? At a deliberately chosen trust boundary where endpoint markings can be validated.
- Why must a priority queue be limited? Unbounded priority traffic can starve other classes.
- Where should an egress queuing policy be applied? At the interface where the bottleneck and output queue actually occur.
Quick self-check
1. Can QoS create more bandwidth?
2. What is the difference between policing and shaping?
3. Where should QoS marking be trusted?
4. Why must a priority queue be limited?
5. Where should an egress queuing policy be applied?
Frequently asked questions
1. Does QoS matter when there is no congestion?
2. Is DSCP end-to-end guaranteed?
3. Why can voice quality be poor despite low average utilisation?
Standards and further reading
- RFC 2475 — An Architecture for Differentiated Services
- RFC 3246 — Expedited Forwarding PHB
- RFC 2597 — Assured Forwarding PHB Group
