QoS on the Enterprise LAN: DSCP and Queue Design for Voice and Video

Is voice breaking up while the uplink looks empty? The DSCP class model, the trust boundary, the priority queue, a Catalyst 9300 example and the order for verifying QoS with counters.
Technical summary
Technology
QoS, DSCP, DiffServ, RFC 4594, Cisco MQC, Catalyst 9000, LLDP-MED
Topic
QoS design for voice and video on a campus network
Level
Intermediate–advanced
Prerequisites
  • VLAN and voice VLAN concepts
  • Basic Cisco IOS XE configuration knowledge
  • Ability to read interface counters
In this article8 sections

Voice is choppy on the phone, faces freeze in video meetings; the network graph, meanwhile, says “the uplink is twenty percent full, no problem.” Most QoS problems on a campus network start here: either it was never configured, or it sits as lines copied from a template that nobody has verified. Voice and video quality isn’t determined by average bandwidth; it’s determined by the few milliseconds when congestion occurs. Below we explain when QoS is needed on the LAN, the DSCP class model, where the trust boundary goes and how the configuration is verified in the field.

When is QoS really needed on the LAN?

QoS doesn’t create bandwidth. It only decides which packet goes first and which is dropped when a queue forms. If there’s no queue, QoS has nothing to do. That’s why the first question is: is there a congestion point in the network? On a campus, these points usually show up in the same places:

  • Speed mismatch: Where traffic from a 10G uplink flows into a 1G access port. A single large file transfer from a server to a client is enough to fill the output queue.
  • Uplink oversubscription: An access switch where 48 × 1G ports share 2×10G uplinks. The average looks low; when the nightly backup, a software rollout or camera recording starts, the instantaneous load spikes.
  • Short traffic bursts (microbursts): An interface that looks lightly used in SNMP’s 5-minute average can overflow its buffer and drop packets at the millisecond scale. The most concrete trace is output drops in the interface counters.
  • Wireless edge: In Wi-Fi the medium is shared. If the DSCP value written on the wired network isn’t mapped correctly to WMM access categories at the access point, the marking ends there.

The decision rule we use in the field is simple: if there are voice or video complaints and output drops are seen on interfaces, QoS is needed. If there are no drops, first isolate the source of latency and jitter; the problem is often Wi-Fi, the WAN line or the end device. For the quality of the line outside the campus, see the WAN line quality and SLA measurement article. Before going into configuration, see the real utilization of uplinks with the uplink utilization and capacity reporting method. The average lies.

Classification, marking and the trust boundary

End-to-end QoS works with three separate jobs:

  1. Classification: Determining which class traffic belongs to (DSCP value, ACL, application recognition).
  2. Marking: Writing the class value into the DSCP field in the IP header. This is done once at the edge of the network; distribution and core queue based on this value, trusting it.
  3. Queuing and scheduling: Putting classes into separate queues on each egress port and determining how much service each queue gets.

The trust boundary is the point where the network starts accepting marking coming from the end device. The rule is clear: don’t trust the marking of a device you don’t control. A user’s computer can mark all its traffic as EF; if the access port accepts this, the priority queue fills up and real voice traffic gets crushed. Set the trust boundary as close to the edge as possible, in practice on the access switch port:

  • Managed IP phones and video endpoints: trust the marking (with conditional trust).
  • User computers: don’t trust. Classify with an ACL on the switch and re-mark, or reduce everything to the default class (DSCP 0).
  • Servers and voice/video infrastructure components (call control server, media gateway): can be trusted on the data center side because you manage them.

DSCP values and a sample class table

DSCP is a 6-bit field in the IP header; it can take 64 values. The basic reference for class names and recommended values is the IETF’s service classes guideline RFC 4594; EF behavior is defined in RFC 3246 and AF groups in RFC 2597. Opening more classes than necessary on a campus makes management harder and the benefit stays limited. Four to six classes are enough for most organizations. The table below is a sample model:

Class DSCP Decimal Queue behavior
Voice (RTP media) EF 46 Priority queue, limited
Interactive video AF41 34 Guaranteed bandwidth share
Call signaling CS3 24 Small guaranteed share
Business-critical data AF21 18 Guaranteed bandwidth share
Default DF (0) 0 Remaining bandwidth
Low-priority bulk CS1 8 Lowest share

RFC 4594 recommends CS5 for signaling; many IP phones and Cisco-based voice environments, however, mark signaling with CS3 by default. Whichever you choose, use the same mapping across the whole network and at the WAN edge. For backup and software distribution, RFC 8622 also defines the “Lower Effort” (LE, DSCP 1) value. Don’t use it without verifying which queue your existing devices put this value in; some older platforms put it in the default queue because they don’t recognize it.

Queuing: the priority queue and buffers

Voice traffic goes to the priority queue (strict priority / LLQ): if there’s a packet in this queue, it’s always sent first. That’s where the risk lies too. If too much traffic enters the priority queue, the other queues starve. That’s why the priority queue is either limited with a policer or a strict trust boundary ensures only real voice traffic gets in. A practical rule that often comes up in Cisco design guides is that the priority queue shouldn’t exceed about a third of link capacity. This isn’t a standard; it’s a starting value.

The other classes get their share with weighted queues (bandwidth remaining). The values here aren’t upper limits but the minimum share during congestion; when there’s no congestion, each class can use the idle bandwidth. You may come across complaints such as “video was capped at thirty percent” that come from mixing up this distinction.

Buffer management is the part skipped most often. On access switches, the buffer is limited memory shared per ASIC. When bursty flows such as video and backup fill the buffer, drops can occur regardless of class. On platforms such as Catalyst 9000, the per-queue buffer ratio and the shared buffer setting (for example qos queue-softmax-multiplier) can be changed. These settings are platform-specific; read the vendor’s QoS configuration guide before touching them. The buffer isn’t infinite.

The voice VLAN and trust with IP phones

In the classic desktop scenario where the computer is connected behind the phone, the access port carries two VLANs: the data VLAN (untagged) and the voice VLAN (802.1Q tagged). The phone learns the voice VLAN information via CDP or LLDP-MED. Separating the voice VLAN doesn’t provide QoS on its own; but it simplifies classification, the DHCP scope and policy enforcement. For the general design of the VLAN structure, the VLAN design in enterprise networks article is the starting point.

Conditional trust: On Cisco switches, the trust device cisco-phone command trusts the marking when a Cisco IP phone is detected on the port via CDP. If the phone is removed and a computer plugged in instead, the trust is automatically removed. With other vendors’ phones, the same result is achieved by applying ACL-based classification to traffic coming from the voice VLAN.

Software clients: Teams, Zoom or Webex run on the computer; their traffic comes from the data VLAN. There are two ways: mark DSCP in the operating system and accept on the switch only marking that comes from expected port ranges, or do the marking directly on the switch based on port ranges. For Microsoft Teams, media port ranges can be defined in the admin center; Microsoft documentation gives source port ranges of 50000–50019 for audio and 50020–50039 for video as examples. On Windows, marking is done with Group Policy or PowerShell:

New-NetQosPolicy -Name "Teams Voice" -AppPathNameMatchCondition "ms-teams.exe" `
  -IPSrcPortStartMatchCondition 50000 -IPSrcPortEndMatchCondition 50019 `
  -DSCPAction 46 -NetworkProfile All

The application path and port ranges may change with the client version. Match the values with your own admin center settings; marking silently disappearing after the client is updated is a familiar fault in the field.

Sample access switch configuration

The example below was prepared with MQC syntax for a Catalyst 9300 running Cisco IOS XE. The ingress policy sets the trust boundary on computer ports; the egress policy defines queuing on uplink and access ports. The percentage values are examples; set them according to your own traffic profile.

ip access-list extended ACL-SOFTPHONE-VOICE
 permit udp any range 50000 50019 any
ip access-list extended ACL-SOFTPHONE-VIDEO
 permit udp any range 50020 50039 any
!
class-map match-any CM-VOICE-IN
 match access-group name ACL-SOFTPHONE-VOICE
class-map match-any CM-VIDEO-IN
 match access-group name ACL-SOFTPHONE-VIDEO
!
policy-map PM-PC-INGRESS
 class CM-VOICE-IN
  set dscp ef
 class CM-VIDEO-IN
  set dscp af41
 class class-default
  set dscp default
!
class-map match-any CM-VOICE
 match dscp ef
class-map match-any CM-VIDEO
 match dscp af41
class-map match-any CM-SIGNALING
 match dscp cs3
class-map match-any CM-BULK
 match dscp cs1
!
policy-map PM-EGRESS
 class CM-VOICE
  priority level 1
  police rate percent 20
 class CM-VIDEO
  bandwidth remaining percent 30
 class CM-SIGNALING
  bandwidth remaining percent 5
 class CM-BULK
  bandwidth remaining percent 5
 class class-default
  bandwidth remaining percent 60
!
interface GigabitEthernet1/0/12
 description USR | Desktop + IP phone
 switchport access vlan 20
 switchport voice vlan 120
 trust device cisco-phone
 service-policy output PM-EGRESS
!
interface GigabitEthernet1/0/30
 description USR | Computer only
 switchport access vlan 20
 service-policy input PM-PC-INGRESS
 service-policy output PM-EGRESS
!
interface TenGigabitEthernet1/1/1
 description UPL | DS01 Te1/0/5
 service-policy output PM-EGRESS

The ACLs look only at the source port range. Another application using the same ports can also get EF; that’s why the egress priority queue is limited with police. On the phone port in the voice VLAN (Gi1/0/12), trust device cisco-phone trusts DSCP only when a Cisco phone is detected via CDP. However, the data VLAN traffic of the computer behind the phone also enters through the same port, and the trust may cover this traffic as well. If you also want to control the computer’s marking, consider adding an ingress policy to the port that re-marks the data VLAN. When conditional trust and service-policy input are used together, verify the platform’s behavior in the Catalyst QoS configuration guide.

There’s no ingress policy on the uplink port; marking coming from the distribution layer is accepted as is because it stays within the trust boundary. Apply the same DSCP–queue mapping at distribution and core too. Otherwise priority applies only in part of the network. The chain breaks at its weakest link.

Verifying QoS and common mistakes

QoS doesn’t count as working just because the configuration was written. It’s verified by measuring marking and queue counters. The order we follow on the console:

  1. Check marking at the edge: In a capture taken on the client side and with SPAN on the uplink, use the Wireshark filter ip.dsfield.dscp == 46 to see that voice packets carry EF. For the capture technique, you can use the network packet analysis with Wireshark guide.
  2. Count policy matches: In the show policy-map interface GigabitEthernet1/0/30 output, verify that the packet counters falling into classes are increasing. If the counter is zero, the ACL isn’t matching.
  3. Monitor queue drops: On Catalyst 9000, show platform hardware fed switch active qos queue stats interface TenGigabitEthernet1/1/1 gives per-queue enqueue and drop counters. Drops should be in the default and bulk queues, and stay at zero in the voice queue.
  4. Measure the application side: In the calling platform’s call quality reports, check that jitter and packet loss stay within an acceptable range under load too.

Mistakes you may encounter:

  • Unconditionally trusting incoming DSCP on computer ports; the priority queue filling with non-voice traffic.
  • Doing QoS at the access layer but not at distribution/core, or using different DSCP–queue mappings between layers.
  • Marking video with EF; opening the priority queue to bursty, high-bandwidth traffic.
  • Not checking the DSCP–WMM mapping on the wireless controller or access point; marking from the wired network getting lost on the Wi-Fi side.
  • A firewall or WAN device resetting the DSCP value; marking never reaching the branch.
  • Thinking QoS is the solution when the source of the problem is uplink capacity or a cabling fault. QoS doesn’t fix a port with CRC errors.

Before starting QoS work, a network performance analysis can be done to clarify the existing drop points and the source of complaints. The class model and trust boundary stay consistent when addressed as part of the segmentation project together with VLAN and routing design.

Frequently asked questions

If we have enough bandwidth, do we need QoS on the LAN?

If there are no output drops on interfaces, QoS’s effect is limited, because QoS only comes into play when a queue forms. But even if average utilization looks low, speed mismatches and traffic bursts cause brief drops. Make the decision based not on average utilization but on drop counters.

Should I trust DSCP marking coming from computers?

No. Reset the incoming marking on computer ports; re-mark only traffic that matches expected applications and port ranges with an ACL on the switch. That way the priority queue doesn’t fill up with non-voice traffic.

Why don’t we mark video traffic with EF?

EF goes to the priority queue, and this queue is served before the others. Video is high-bandwidth and bursty; if it’s put in the priority queue, it can starve the other classes. For interactive video, giving it a guaranteed bandwidth share with AF41 gives a more balanced result.

How do I prove QoS is working?

Show with a packet capture that DSCP values are carried correctly, with policy-map counters that there are matches to classes, and with queue statistics that drops occur not in the voice queue but in low-priority queues. Run a call test under load and record jitter and loss values.

Support

Contact us for support or a discovery session on this topic.

Contact form
[email protected]

Eretek Network

We build switching, routing, cabling and wireless networks with VLAN segmentation and redundant uplinks, and provide regular maintenance and fault support.

Knowledge Center

Other articles