QoS basics: keeping voice and video clear

Why QoS?

Most of the time links are not full and every packet goes straight through. But at busy moments, such as a big backup, a software update on 100 laptops, or month-end reports, packets wait in a queue at the router. Email and file downloads do not mind waiting a little. A voice call or video meeting does: the words arrive late, choppy or not at all.

Four things describe how a network treats traffic. Bandwidth is how much data per second a link can carry. Delay (latency) is how long a packet takes to arrive. Jitter is the variation in delay from packet to packet. Loss is the share of packets that never arrive. For good voice quality, Cisco's common targets are a one-way delay of 150 ms or less, jitter of 30 ms or less, and loss of 1% or less. A voice call needs little bandwidth (roughly 30 to 90 kbps per call depending on the codec), but it needs it on time.

Step 1: classification and marking

Classification means recognising traffic: this packet is a voice call, that one is a backup. You can classify by ACL, by application recognition (Cisco NBAR), or by trusting a marking already on the packet. Marking means writing a value into the packet header so that every later device can classify it with one quick look.

• CoS (Class of Service): 3 bits in the 802.1Q VLAN tag, values 0 to 7. It exists only on trunk links (and the voice VLAN), so it is lost when a router removes the Ethernet header. Voice is usually CoS 5.
• DSCP (Differentiated Services Code Point): 6 bits in the IP header's ToS byte, values 0 to 63. It stays with the packet end to end across routers. The older IP Precedence used only the first 3 bits.

AF (Assured Forwarding) names have two digits: AFxy. x is the class (1 to 4; higher classes usually get better treatment) and y is the drop precedence (1 to 3; higher is dropped first inside its class). The DSCP value is 8x + 2y, so AF41 = 32 + 2 = 34 and AF13 = 8 + 6 = 14. CS (class selector) values are multiples of 8: CSn = 8n, matching the old IP Precedence values.

The trust boundary

If every PC could mark its own traffic EF, a student could make a game download "voice". The trust boundary is the point where the network starts believing markings. A common design trusts the Cisco IP phone on the access switch port: the phone marks its voice EF and CoS 5, while traffic from the PC plugged into the phone is re-marked to 0. Everything beyond the access switch then trusts the markings.

Step 2: queuing

Queuing decides the order in which waiting packets leave an interface when it is congested. FIFO (first in, first out) has one queue and no favourites. CBWFQ (class-based weighted fair queuing) gives each class its own queue and a guaranteed minimum share of bandwidth. LLQ (low latency queuing) adds one strict priority queue to CBWFQ: voice packets in it always go first, but the priority queue is limited to a set amount so it cannot starve the rest.

class-map match-any VOICE
 match dscp ef
class-map match-any VIDEO
 match dscp af41
policy-map WAN-EDGE
 class VOICE
  priority percent 20
 class VIDEO
  bandwidth percent 30
 class class-default
  fair-queue
interface GigabitEthernet0/1
 service-policy output WAN-EDGE

Step 3: congestion avoidance

When a queue is completely full, new packets are simply dropped: tail drop. Many TCP connections then slow down together, then speed up together, so the link swings between overloaded and half empty (TCP global synchronisation). RED (random early detection) starts dropping a few random packets before the queue is full, so TCP flows slow down at different times. WRED (weighted RED) drops lower-priority markings first, for example AF13 before AF11.

Step 4: policing and shaping

Both limit traffic to a rate. Policing drops (or re-marks) anything over the rate at once, with no extra delay; ISPs police what customers send into their network. Shaping holds the extra packets in a buffer and sends them a little later, smoothing the traffic to the rate; customers shape their outgoing traffic to the rate they bought, for example 200 Mbps on a 1 Gbps port, so the ISP's policer does not drop it. Shaping adds some delay; policing adds loss.