Designing a Simple QoS Policy for a Small Network
Search for a QoS policy template and you will find twelve-class designs with names like "multimedia streaming" and "transactional data," each with its own DSCP value, bandwidth percentage, and drop profile. Those designs exist for large networks with teams to maintain them. Copy one onto a fifty-person office and you get twelve classes nobody can explain and marking rules nobody owns. You also get a phone system that still breaks up during backups because the one class that mattered was configured wrong.
This guide takes the opposite approach. You will design a policy with three or four classes and find the handful of links where it actually matters. You will apply it at the WAN edge first because that is where the return is. You will set marking and trust so the classes see real traffic. And you will roll it out in an order that cannot take the network down. At the end you will have a one-page policy document and a checklist. If the words classify, mark, queue, and shape are not yet familiar, QoS in Plain Language explains them in twenty minutes.
In this guide
- Start with three or four classes, not twelve
- Find your actual congestion points first
- Apply policy at the WAN edge before anywhere else
- Consistent marking and trust at the access layer
- Wi-Fi, in one paragraph
- The sample policy, on one page
- What to document
- Rollout order: monitor first, then enforce
- Where to go next
Start with three or four classes, not twelve
A class is a group of traffic you have decided to treat the same way. Every class you create is a decision you must classify, mark, allocate bandwidth to, monitor, and be able to explain to the person whose traffic landed in it. Four is the most a small network needs, and three is often enough.
- Real-time. Voice audio and interactive video. Small, steady streams that are ruined by delay and cannot be retransmitted. This is the strict-priority class, and it is the reason you are doing any of this. If video meetings are a large share of your day, they can justify a class of their own; prioritizing video traffic covers when that split is worth it.
- Business-critical. The one or two applications the business actually runs on: the ERP or line-of-business system, remote desktop sessions, and usually call signaling. Interactive, sensitive to delay, but tolerant of a little loss. This class gets a guaranteed share, not priority.
- Best effort. Everything else. Web, email, file shares, printing. This is the default class, and most of your traffic lives here on purpose. Never police it and never starve it; it is the class users notice first when something is wrong.
- Scavenger. Traffic that should lose first when the link is tight: backups, replication, operating system updates, cloud sync clients. It gets a tiny guaranteed minimum so it never stalls completely and borrows freely whenever the link is idle, which is most of the night.
A useful test for whether something deserves a class: can you name the traffic and the person who will complain if it is treated badly? Can you name the person who will complain if it is treated too well? If any answer is "not sure," it belongs in best effort. Resist the urge to create a class for something because a template had one. Call signaling is the usual judgment call. It is small and it matters, but it tolerates a little delay. So business-critical suits it better than the priority queue, where it would compete with the audio it sets up.
Find your actual congestion points first
QoS does nothing on a link that never fills. So before writing a single class, find the links where queues actually form, because those are the only places the policy will earn its keep. In a small network the list is short and predictable, but confirm it with data rather than assumption:
- Interface utilization at short intervals. Look at the WAN interfaces and every uplink, and look at peaks rather than averages. A five-minute average hides the ten-second bursts that ruin calls.
- Output discards. The standard interface counters include a "discarded on output" value that most devices increment when an output queue overflows. It is the cheapest queue-drop signal you have, and every monitoring tool can collect it.
- Complaint timing. If call quality tickets cluster around the backup window or the morning login rush, you have found congestion without a single graph.
- Speed steps on the port list. Every place a faster link feeds a slower one is a candidate. Examples include the gigabit LAN into the WAN circuit, or the office into the VPN tunnel. Others include wired into wireless, or a fast server into a slow client segment. Tunnels deserve a second look, because the encrypting device often has a throughput ceiling well below the circuit it sits on. That ceiling is a congestion point of its own.
For most small networks the ranking comes out the same way. The WAN circuit's outbound direction comes first, the site-to-site tunnel second, and Wi-Fi third. Then perhaps comes one uplink to a busy server closet. Switch uplinks inside a small office are usually idle by comparison; monitoring uplinks shows how to confirm that rather than guess. And if the data says nothing congests at all, be honest about it. In that case, apply the policy at the WAN edge anyway, because it is cheap and it will matter the day the link fills. Skip everything else.
Apply policy at the WAN edge before anywhere else
The WAN edge is where the network's fastest links meet its slowest, on one device, on one interface, in one direction you fully control. Nowhere else offers that combination, which is why a policy applied here alone delivers most of the benefit of a full deployment. The design has four parts, and the order matters.
- Shape to the real circuit rate. The router's WAN port is faster than the circuit behind it, so without a shaper the queue forms at the provider, where your policy does not exist. Shape outbound to slightly under the rate the circuit actually delivers, measured with a throughput test rather than read off the contract. A few percent under is the usual margin, enough to keep the queue on your side without wasting capacity.
- Give real-time strict priority with a cap. The priority queue is served first, always. So it must be limited to a fraction of the shaped rate or it can starve everything else. Around a quarter to a third of the link is a common ceiling.
- Guarantee shares for the rest. Business-critical gets a meaningful minimum, best effort gets the largest share, scavenger gets a sliver.
- Classify by mark, with a fallback. Match DSCP first, but add address or port rules for traffic whose marks may have been lost on the way in, especially anything arriving from the far side of the internet.
Written in a neutral pseudo-configuration, the whole thing fits on a screen. Your platform's syntax will differ; the structure will not.
# Classes: recognize traffic by its mark, with an address fallback
class REALTIME match dscp EF
class BUSINESS match dscp AF31, AF41
match dst-net 10.50.0.0/24 # ERP servers, in case marks were lost
class SCAVENGER match dscp CS1
# anything unmatched is BEST-EFFORT, the default class
# Outbound policy for the WAN interface
policy WAN-EDGE-OUT
shape 95 Mbps # just under the measured circuit rate
queue REALTIME priority, limit 25% # strict priority, capped
queue BUSINESS guarantee 30%
queue SCAVENGER guarantee 5%
queue BEST-EFFORT guarantee remainder, fair-queue
interface wan0
apply policy WAN-EDGE-OUT outbound
Two notes on what is missing. There is no inbound policy, because inbound congestion happens on the provider's side, where your queues do not exist. Some devices can shape inbound traffic anyway, and it helps a little, because dropping a few packets after they arrive makes remote senders slow down. Treat it as a refinement for later, not the place you start. And there is only one site shown. A private WAN or a site-to-site tunnel needs the mirror image of this policy at the far end, or only one direction of every call is protected. The specific bandwidth per call and the delay, jitter, and loss targets decide how big the real-time cap must be. Those numbers are laid out in QoS for real-time traffic. Size the priority queue from those numbers rather than from a guess.
Consistent marking and trust at the access layer
The WAN-edge policy classifies by DSCP, so its classes only fill if marks arrive intact. That is the access layer's job, and it has two halves: decide where each class gets its mark, and decide which ports are believed. For each class, pick one source of truth:
| Class | Who marks it | Where | Why there |
|---|---|---|---|
| Real-time (phones) | The phone itself | At the endpoint; the switch trusts the phone port | Phones mark reliably and know audio from signaling |
| Real-time (conferencing on PCs) | The application, if the operating system allows it; otherwise the WAN router by port and address | Endpoint or WAN edge | PC marks are unreliable; do not build the policy on them |
| Business-critical | The switch, by server port and VLAN; the router by destination address for the client-to-server direction | Access switch and WAN edge | Servers do not mark themselves; the switch does not care what they think |
| Scavenger | The switch, on the backup server's port | Access switch | Backup software almost never marks itself |
| Best effort | Nobody — untrusted ports rewrite to 0 | Every user-facing port | The default needs no mark, and the rewrite stops users self-promoting |
Trust follows from the table. Ports facing phones, ports facing servers you have configured to be marked, and every uplink between your own switches trust DSCP. Ports facing user PCs, guests, and printers do not, and rewrite whatever arrives to 0. (The printer does not need priority. The printer has never needed anything except attention.) The rule that catches people is the uplinks. A mark that survives the access switch and is zeroed at the distribution switch is just as dead. So trust must be set on both ends of every link between the marking point and the WAN router.
Consistency is the other half. Every device on the path must agree that EF means real-time and CS1 means scavenger. Write the mapping once, apply the same values at every site, and never let one office use AF31 for business-critical while another uses AF21. When marks go missing anyway, and they will, DSCP Markings in the Wild walks the devices that strip them.
Wi-Fi, in one paragraph
Access points do not queue by DSCP. They translate it into one of four wireless categories (voice, video, best effort, background) and schedule airtime by category. Your whole Wi-Fi design is therefore a mapping check. Turn on the wireless multimedia QoS feature if it is off. Confirm that EF lands in the voice category, AF41 in video, and CS1 in background. Confirm the access point rewrites DSCP sensibly when frames come back onto the wire. Accept that upstream marking is largely up to the client device. Then stop, because on Wi-Fi, channel planning, client density, and interference decide call quality far more than any mapping table, and those are a different project.
The sample policy, on one page
Here is the entire design as the single table you will hand to whoever maintains it after you. Percentages are of the shaped WAN rate and are starting points to adjust from your own per-class graphs, not laws.
| Class | Contents | DSCP | Marked at | WAN-edge treatment | Healthy looks like |
|---|---|---|---|---|---|
| Real-time | Voice audio, interactive video | EF (46), video AF41 (34) | Phone; app or WAN router for PCs | Strict priority, capped at 25% | Zero drops, always |
| Business-critical | ERP, remote desktop, call signaling | AF31 (26); signaling CS3 (24) | Access switch and WAN router | Guaranteed 30% | Rare drops; none sustained |
| Best effort | Everything else | 0 | Nowhere (untrusted ports rewrite to 0) | Guaranteed remainder, fair-queued | Drops under load are expected |
| Scavenger | Backups, replication, updates, cloud sync | CS1 (8) | Access switch, server ports | Guaranteed 5%, borrows when idle | Drops during the busy day are the design working |
Tip: leave routing-protocol and management traffic out of your classes entirely. Most routers protect their own control traffic automatically. And the worst thing you can do is catch it in a scavenger rule and watch a routing adjacency flap under load. If a device does not protect it by default, give it a small class of its own rather than folding it into business-critical.
What to document
A QoS policy that lives only in device configurations is not a policy; it is an archaeological site waiting for your successor. The document is short, and it is the thing that makes the policy survivable. Include:
- The class table above, with the DSCP values, exactly as deployed.
- Where policy is applied. Each device, interface, direction, and the shaped rate, along with how the circuit rate was measured and when.
- Trust boundaries. Which ports trust DSCP, which mark, and which rewrite to 0, ideally as a rule ("phone and server ports trust; user ports do not") plus the exceptions.
- Fallback classification rules. Every address- or port-based match, with the reason it exists and the address it depends on, so it can be retired when the server moves.
- The provider's honored classes, if you have a private WAN, and how yours map to them.
- An owner and a review date. Address rules go stale; circuits get upgraded; the shaped rate must follow.
- The test procedure and the last results, so the next person can prove it still works instead of trusting it does.
- A change log. One line per change, in order, saying what was altered and why. When call quality changes, the first question is always "what changed," and this is what answers it.
Rollout order: monitor first, then enforce
The most common way QoS projects break a network is by doing everything at once. Each step below changes one thing, is observable before the next step starts, and can be undone in a minute by removing a policy from an interface. Work through them in order, in a maintenance window for the enforcing steps, with someone ready to make a test call.
- 1. Baseline. Collect a week of WAN and uplink utilization, output discards, and a note of when calls are bad. You need a "before" to compare against.
- 2. Mark and trust at the access layer. Set trust on phone, server, and uplink ports; add the switch marking rules for business-critical and scavenger. Nothing queues yet, so this cannot slow anyone down.
- 3. Verify marks arrive. Capture on the LAN side of the WAN router and confirm each class's traffic carries the expected DSCP. Zero EF in the capture means step 2 is not finished.
- 4. Classify in monitor-only mode. Apply the class definitions at the WAN edge with no shaping or queuing actions, just counters. Watch the per-class match counts for a day. Every class should count roughly the traffic you expect; a class at zero means the marks are not reaching this device.
- 5. Add the shaper alone. Enable the shaper at the measured circuit rate with no per-class queuing. Confirm throughput is unchanged and nothing stalls. If it does, the shaped rate is wrong, not the classes.
- 6. Enable queuing. Turn on the priority queue and guaranteed shares. Make test calls during the busiest hour. Expected: best-effort and scavenger drops appear under load, the priority class shows none.
- 7. Prove it under load and record the result. Generate load, measure each class, and keep the evidence. That is its own guide: Proving QoS Works End to End.
- 8. Repeat outward. The far site's WAN edge, then the tunnel, then the Wi-Fi mapping, then any uplink the baseline showed congesting. One device per change.
A few rules keep the rollout boring, which is the goal. Never police best effort; shape and share, but do not drop the default class on purpose. Never put "unknown" traffic in scavenger; unknown belongs in best effort until you know what it is. Keep the priority cap even when it seems generous, because the day it is exceeded is the day the priority queue starts dropping voice. Have the rollback written down before you start. On most devices it is a single command that removes the policy from the interface. Knowing that makes the enforcing steps far less nerve-wracking. Change one device per window, and announce the window, so that a complaint the next morning can be tied to a step instead of to a mystery. And after the policy is live, watch the per-class counters for a week — Monitoring Queue Drops explains what each counter means and what to alert on.
Where to go next
A small-network QoS policy has four classes, one shaped WAN interface, and a trust boundary at the access layer. It has a one-page document and a rollout that changes one thing at a time. Start at the WAN edge because that is where the line forms. Mark at the source you can trust. And keep the design small enough that you can explain every class to the person it affects.
From here, the natural next steps are watching the policy work through its counters and then proving it under deliberate load. Both are covered in this series, along with the plain-language mechanics and the hunt for stripped markings. Everything is gathered on the QoS: Designing, Verifying & Monitoring It pillar page.
Put this into practice with SysPeep Monitor
A QoS policy is a promise about how your links behave under load. And the only way to know whether the promise holds is to watch the links. SysPeep Monitor polls your WAN routers and switches over SNMP. It graphs utilization and discards per interface, and alerts when a link you shaped starts dropping — from one web console. You can install the free trial on a Windows machine today.
Download Free TrialLearn About SysPeep