← Interview Prep

QoS — Theory

Quality of Service for interviews: IntServ vs DiffServ, classification & marking (DSCP/802.1p/MPLS TC), PHBs (EF/AF/CS), queuing (LLQ/CBWFQ), WRED/ECN, policing vs shaping, and MPLS DiffServ tunneling modes.

QoS is how a congested link decides whose packets wait and whose get dropped. On an uncongested link it does nothing — queues are empty. The whole discipline is about managing the queue that forms when demand briefly exceeds the egress rate: classify traffic, mark it, and give each class the latency, loss, and bandwidth it needs.

The three models

Model	Idea	Scaling
Best-effort	No differentiation; FIFO.	Trivial, no guarantees.
IntServ	Per-flow reservations signalled with RSVP.	Per-flow state — doesn't scale in the core.
DiffServ	Mark packets into a few classes; each router applies a per-hop behavior.	Stateless per-flow — the model everyone actually deploys.

Classification & marking

Classify once at the edge, mark the header, and trust the mark downstream. The mark lives at different layers:

Marking	Where	Bits
DSCP	IPv4 ToS / IPv6 Traffic Class	6 bits (64 values) — the primary L3 marker
IP Precedence	Top 3 bits of the old ToS byte	Legacy; DSCP superset
802.1p (PCP)	802.1Q VLAN tag	3 bits (8 CoS) — L2 only
MPLS TC (EXP)	MPLS label	3 bits — QoS across an MPLS core

Trust boundary: decide where you believe incoming marks. Untrusted access ports get re-marked (or set to 0); trusted uplinks keep the DSCP. This stops a host from marking its own traffic EF and stealing the priority queue.

Per-Hop Behaviors (the standard DSCP classes)

PHB	DSCP	For
EF (Expedited Forwarding)	46	Voice / low-latency — strict priority, policed.
AFxy (Assured Forwarding)	AF11…AF43	4 classes (x) × 3 drop precedences (y) — video, business apps.
CSx (Class Selector)	CS0–CS7	IP-Prec-compatible; CS6/7 for control plane/routing.
BE / Default	0	Everything else.

AF drop precedence matters at congestion: within a class, higher-drop-precedence packets (e.g. AF13) are discarded before lower (AF11) — how a policer marks "in-contract" vs "out-of-contract" traffic.

Queuing & scheduling

Each egress class gets a queue; the scheduler decides service order:

FIFO — one queue, no differentiation.
PQ (Priority Queuing) — strict priority; can starve lower classes.
WRR / DWRR / WFQ / CBWFQ — weighted sharing so each class gets a guaranteed slice of bandwidth without starvation.
LLQ (Low-Latency Queuing) = a policed strict-priority queue (for EF/voice) on top of CBWFQ — the common enterprise design: voice jumps the line but can't starve the rest.

Congestion avoidance

Tail drop: queue full → drop new arrivals. Causes global TCP synchronization (many flows back off together, sawtooth).
WRED (Weighted Random Early Detection): start dropping probabilistically before the queue is full, worse-drop-precedence first — desynchronizes TCP and keeps queues shallower.
ECN: instead of dropping, mark the packet (CE) so the receiver echoes it and the sender backs off without loss — increasingly the DC standard (DCTCP).

Policing vs shaping

	Policing	Shaping
Excess traffic	Dropped (or re-marked)	Buffered & sent later
Effect on rate	Hard cap, bursty output	Smooths to the rate
Latency	None added	Adds delay/jitter (queue)
Direction	In or out	Egress

Both use a token bucket: rate = CIR, burst = Bc (+ Be / PIR for two-rate). Tokens accumulate at CIR; a packet is conforming if tokens are available.

QoS across an MPLS core (tunneling modes)

When IP is tunneled in MPLS, how does the label's TC relate to the inner DSCP on push/pop?

Uniform: one QoS domain — DSCP↔TC copied both ways; core re-marks propagate to IP.
Pipe: provider QoS hidden from the customer — egress uses the tunnel's TC for the final PHB but forwards the original DSCP untouched.
Short-pipe: like pipe, but egress uses the customer DSCP for the final PHB.

The low-latency/trading angle

For HFT, "QoS" often means the opposite of buffering: you want minimum queue depth. Strict priority for the trading class, tiny buffers, cut-through switching, and no shaping — because a shaped queue is added latency. See Low-Latency & Trading.

Likely interview questions

IntServ vs DiffServ — why doesn't the core do per-flow reservations?
Policing vs shaping: which drops, which buffers, and what does each do to latency?
What is a trust boundary and why does it matter? (hosts faking EF)
EF vs AF vs CS — what's each for, and what does AF drop precedence do?
Why WRED/ECN instead of tail drop? (global TCP synchronization)
What is LLQ and why not just strict PQ everywhere? (starvation)
How is QoS carried across an MPLS core, and what are pipe vs uniform modes?

Related: Low-Latency & Trading · MPLS Theory.