QoS (Quality of Service)

How QoS works, in four stages

Every QoS policy, on any vendor’s equipment, does the same four things in the same order. The vendor menus differ; the mechanism does not.

Classify

Identify which packets are voice. Best done by the phone or the switch port, not by guessing at port numbers, because SIP signalling and RTP media use different ports and RTP ranges vary by platform.

Mark

Write a value into the packet header so every device downstream can act without re-inspecting it. Voice media gets DSCP EF, decimal 46, defined by RFC 3246 and recommended for the Telephony class by RFC 4594.

Queue

Put marked packets in a strict priority queue on the interface where congestion happens. That is almost always the WAN uplink, not the office switch.

Police

Cap the priority queue. Cisco’s own guidance is blunt about why: without a policer on the priority class, lower priority queues can be starved of bandwidth entirely. Size the cap to your real concurrent call count.

Sizing the cap is arithmetic, not judgement. A G.711 call consumes about 87 kbps on Ethernet once headers are counted, so twelve concurrent calls need roughly 1 Mbps of protected capacity. Reserve for the calls you actually place at the busiest hour, not for every extension on the account.

Which markings to use

Three values cover a business voice deployment. Set them once and make every device agree.

The choice between CS3 and CS5 matters less than consistency. A network where the handsets mark CS5, the switch trusts CS3 and the router matches on port numbers is a network with no working QoS at all, however many policies are configured.

Where your QoS stops working

Markings are honoured inside a network that agreed to honour them, and nowhere else. RFC 2474 is explicit: boundary nodes must ensure traffic entering a domain carries codepoints appropriate to that domain, remarking the traffic with new codepoint values if necessary, and translations at those boundaries are a matter for the service level agreement between provider and customer.

In practice that means your EF marking is respected on your LAN and on your own uplink queue. Once the packet crosses into your ISP’s network it is honoured only if the ISP has agreed to honour it, and across the wider internet it is commonly reset. Nobody in Cape Town is prioritising your voice packets on your behalf.

There is a second limit that catches most offices out. You can only queue traffic you are sending. Inbound congestion is decided by whoever is upstream of your bottleneck, so if the fault is a saturated download path, the fix is either more capacity or an ISP that shapes on your behalf, not a router policy. Since a call is symmetrical, both directions have to be sound before the call is.

Setting it up on a South African office network

1. Put phones on their own VLAN

   A voice VLAN gives you a clean way to classify traffic by where it came from rather than by inspecting it, and it keeps handset broadcast traffic away from the rest of the office.

2. Mark at the edge and trust it

   Let the handset or the access switch apply DSCP EF and 802.1p 5, then configure the uplink to trust those marks. Marking twice, in two places, with two different values is the most common self-inflicted QoS fault.

3. Apply the priority queue on the WAN interface

   Congestion happens where a fast network meets a slower line. A gigabit switch is not the bottleneck; the uplink is. Policies on the wrong interface do nothing at all.

4. Size and cap the queue

   Count your genuine peak concurrent calls, multiply by the per-call figure for your codec, and set the cap there. Leaving the priority class uncapped simply moves the problem to your data traffic.

5. Ask your ISP what happens next

   Ask, in writing, whether the ISP honours DSCP on their network and what they do with EF-marked traffic at the handover. A yes is worth having. A vague answer tells you your QoS ends at your own router.

6. Test at the busiest hour

   QoS that has never been tested under load has not been tested. Start a large upload, place a call, and listen. If the call survives, the policy works.