QoS, Latency, Jitter, and Packet Loss

Bandwidth is capacity, not performance

A 1 Gbps link describes capacity. It does not reveal short bursts, time spent in switch buffers, dropped packets, or whether a critical packet is waiting behind a large transfer.

Several encoders can transmit at nearly the same moment and briefly exceed an outgoing interface's line rate. Packets collect in its buffer and are discarded if the buffer fills. A five-minute utilization graph may still look comfortable. This short-duration congestion is called a microburst.

Oversubscription occurs when the possible incoming load exceeds an outgoing link. Document endpoint speeds, stream rates, simultaneous flows, traffic paths, uplink capacity, and headroom—not just the switch port count.

Latency is the total travel time

Latency includes processing, propagation, serialization, and queuing delay. A device must inspect a packet; the interface must place every bit on the wire; the signal must travel; and the packet may have to wait behind other traffic.

A 1,500-byte packet contains roughly 12,000 bits before additional Ethernet overhead. Serialization alone takes about 120 microseconds at 100 Mbps, 12 microseconds at 1 Gbps, and 1.2 microseconds at 10 Gbps. Within a building, variable queuing and device processing often matter more than propagation distance.

Jitter is inconsistent delay

Jitter is variation in packet arrival time. Real-time receivers use a buffer to release unevenly arriving packets at a steady rate. A larger buffer tolerates more variation but adds latency; a smaller buffer reduces latency but gives the network less room for error.

That is why a lower Dante latency setting is not automatically better. The correct setting is one the complete path can meet consistently, including every switch hop and congested interface.

Packet loss misses the deadline

Packets may be lost because of congested queues, bad links, failing optics, wireless interference, endpoint overload, multicast errors, or loops. TCP can request missing data again, but real-time UDP media often cannot wait for retransmission.

The receiver may conceal the loss, repeat old information, mute, freeze, show artifacts, or drop the frame. Increasing error, discard, queue-drop, or late-packet counters is evidence; 'it feels networky' remains more of a spiritual diagnosis.

QoS is a complete handling chain

QoS classifies traffic, marks it, decides where markings are trusted, maps each class into a hardware queue, and schedules those queues during contention. If the interface is idle, priority and ordinary packets both leave immediately.

DSCP is a six-bit Layer 3 marking in the IP header. Dante commonly uses DSCP 56 for time-critical PTP event traffic and 46 for real-time audio, but every platform's current guidance must be verified. PCP is the three-bit priority field in an 802.1Q tag and exists only while that Ethernet frame is tagged.

A marking is only a label. Every switch in the path must trust or intentionally rewrite it, map it to the intended queue, and schedule that queue correctly. Strict priority is useful for small amounts of critical traffic, but misplacing a large stream there can starve everything below it.

• Classify the traffic.
• Mark it with DSCP or PCP where appropriate.
• Define the trust boundary.
• Map markings into hardware queues.
• Schedule and monitor the queues under load.

QoS cannot repeal math

If two sustained 700 Mbps streams must cross a 1 Gbps uplink, QoS cannot fit 1.4 Gbps through the pipe. It may protect clock and control packets, but some video must queue or drop.

The real correction may be a faster uplink, lower stream rates, fewer simultaneous flows, different source placement, or an appropriate link-aggregation design. QoS decides who goes first during contention; it does not manufacture capacity.