Fine-tuning QoS Shaping and Hierarchical QoS (HQoS) FAQ: Expert Answers to Technical & Deployment Questions

Fine-tuning QoS Shaping and Hierarchical QoS (HQoS) FAQ: Expert Answers to Technical & Deployment Questions

Overview & Thematic Scope

This FAQ covers the practical fine-tuning of QoS shaping and hierarchical QoS (HQoS) on carrier Ethernet and enterprise routing platforms. It targets the questions network engineers and procurement teams actually raise: how shapers differ from policers, when hierarchical scheduling is worth the overhead, how to split a port’s bandwidth across subscribers, and how to prove the configuration works in production. Answers reflect typical merchant-silicon and dedicated-ASIC behaviour, so always validate timers, queue counts and token-bucket granularity against your specific platform’s release notes.

Fine-tuning QoS Shaping and Hierarchical QoS (HQoS) FAQ: Expert Answers to Technical & Deployment Questions details

Frequently Asked Questions

Q1: What is the difference between QoS shaping and policing, and which should I deploy?
Shaping buffers excess traffic and releases it at the configured rate, while policing drops or remarks excess traffic immediately. Shaping smooths bursts and is preferred on egress toward a slower WAN or subscriber link; policing is used on ingress to enforce a contracted rate and protect downstream resources. Most designs pair ingress policing with egress shaping for a complete per-subscriber SLA.
Q2: What is hierarchical QoS (HQoS) and when is it actually required?
HQoS applies multiple levels of shaping and scheduling — typically port, then group or subscriber, then class of service — so bandwidth guarantees are enforced at every tier. It is required when one physical port carries many subscribers or service instances and you must guarantee each one a share of the port without starving others. Single-subscriber uplinks and simple campus links rarely justify the added configuration and CPU cost.
Q3: How do I split a 10G port across multiple subscribers or service groups with HQoS?
Configure a parent shaper at the port level, child shapers per subscriber or VLAN group, and a class scheduler under each child for voice, business and best-effort traffic. The parent shaper should never exceed the physical line rate minus Layer 2 overhead, and the sum of child committed rates must not exceed the parent rate. Reserve strict-priority queues for real-time classes capped at roughly 10–30% of the parent rate to avoid starving other traffic.
Q4: What oversubscription ratio is safe between the sum of child shapers and the parent port?
A 2:1 to 4:1 ratio between aggregated child committed rates and the parent port is typical for residential or guest traffic, while business SLAs usually stay at 1:1 or below 2:1. The safe value depends on burst tolerance and the platform’s buffer depth per queue, not just the arithmetic ratio. Model peak-hour concurrency before committing to a ratio, and monitor queue drops per class after go-live.
Q5: Why does my shaper show the configured rate but still drop traffic, and how do I verify it?
Drops usually mean the child or class queue, not the shaper itself, is oversubscribed or misconfigured relative to the parent. Verify with per-queue statistics, shaper token and drop counters, and interface output rates against the configured CIR and PIR. Also check Layer 2 overhead assumptions, because shaper rates calculated on payload alone will under-deliver on small-packet traffic.
Q6: How do I tune QoS scheduling to reduce latency for voice and real-time traffic?
Place real-time traffic in a strict-priority or low-latency queue, cap that queue so it cannot monopolize the port, and keep the remaining classes on weighted scheduling. Packet serialization on the egress port is often the dominant latency component, so prioritization only helps if the priority queue itself is not congested. Measure one-way delay per class with timestamped probes rather than relying on interface counters alone.
Q7: How many queues and HQoS levels can typical hardware support?
Most access and aggregation platforms support 8 hardware egress queues per port and 2–3 HQoS levels, with larger core chassis offering more queues but shallower hierarchy per port. The practical limit is often memory for per-queue buffers rather than queue count, and merchant-silicon platforms share a buffer pool across ports. Confirm both the queue count and the per-queue burst depth on the datasheet before designing multi-level subscriber hierarchies.
Q8: What are the most common HQoS misconfigurations and how do I troubleshoot them?
The most common errors are child committed rates summing above the parent, a strict-priority class left uncapped, and QoS policies applied to the wrong direction. Troubleshoot by inspecting policy attachment points, verifying per-level shaper statistics, and confirming trust boundaries so DSCP markings are not rewritten upstream. A short pre-deployment test with synthetic traffic at each hierarchy level catches nearly all of these before production.