LTE QoS Model (QCI / ARP / GBR)
The QoS framework: the standardised QCI table, ARP, GBR vs non-GBR, and how it maps to bearers.
In LTE, quality of service is a property of a bearer, not of an individual packet. Every EPS bearer carries exactly one QCI, and that single number tells the whole system — from the P-GW down to the eNB scheduler — how to treat the traffic that rides inside it. This is the fundamental contrast with 5G, where QoS is per-flow. The framework is defined in TS 23.401 (EPS architecture and bearer procedures) and TS 23.203 (policy and charging control).
Introduction
QoS in LTE is the machinery that lets one network carry a low-latency voice call, a bulk file download and IMS signalling at the same time, and give each the treatment it needs. The whole framework is built on one design choice: QoS is attached to the bearer, and every bearer carries a single scalar label, the QCI, that stands in for a whole set of standardised delay, loss and priority targets.
This model touches every part of the system. Policy nodes (PCRF/PCEF) decide the QoS and which flow gets which bearer; the core marks and forwards per bearer; and the eNB scheduler enforces it every millisecond by choosing which bearer's logical channel to serve on the air interface. When you debug "why is my voice choppy" or "why did my video not get its rate", you are almost always tracing a value — a QCI, an ARP, a GBR — from where policy set it to where the scheduler acted on it.
It matters because QoS is where an operator's commercial and regulatory intent becomes concrete radio behaviour: subscriber tiers, VoLTE guarantees, emergency-call pre-emption and fair-use caps are all expressed through the parameters on this page.
On this page
Why LTE QoS is needed
In plain words: think of an airline. A QCI is the cabin class printed on your ticket — it fixes, in advance, how you are treated in the air (meal, legroom, priority of service) using one label everyone understands. ARP is separate: it is the boarding and bumping policy — who gets a seat when the flight is oversold, and whose seat can be given away. And GBR/AMBR are the baggage rules: a guaranteed allowance for premium passengers, versus a shared, capped allowance for everyone in economy.
Concretely, LTE needs a QoS framework because a single radio link is shared and finite, and different applications have contradictory needs. Voice wants low, bounded delay and can shrug off the odd lost frame; a download wants raw throughput and cannot tolerate loss; signalling wants to be served first but uses almost no bandwidth. Without a QoS model the scheduler would treat all of these identically and all of them would suffer under load. By stamping each bearer with a QCI (and an ARP, and where relevant a GBR), LTE lets every node apply the right treatment consistently, decide what to admit when capacity is scarce, and cap what best-effort traffic may consume — all without inspecting individual packets.
A per-bearer QoS model: one QCI (standardised delay/loss/priority), one ARP (admission/pre-emption), and bit-rate parameters (GBR/MBR for GBR bearers, APN-AMBR/UE-AMBR for non-GBR).
To give contradictory traffic types consistent, predictable treatment on a shared, finite radio, and to decide what to admit and what to cap under congestion.
The PCRF decides policy and pushes a PCC rule to the PCEF in the P-GW, which activates a bearer with the right QCI/ARP/GBR and installs a TFT; the eNB scheduler then enforces it per logical channel.
The EPS bearer — the unit of QoS in LTE
In the EPS, the EPS bearer is the finest granularity of QoS differentiation. Every IP packet that receives the same forwarding treatment — the same scheduling, queue management and rate policing — travels in the same bearer, and the mapping is fixed at bearer setup time. A UE gets exactly one default bearer per PDN connection (always non-GBR, brought up when the PDN connection is established and kept alive for its entire life), and it may get additional dedicated bearers layered on top for traffic that needs its own treatment — for example a GBR bearer for a VoLTE call, or a separate non-GBR bearer with a better QCI for premium video.
An EPS bearer is a logical pipe between the UE and the P-GW with one agreed QoS. It is characterised by a QCI, an ARP, and — for GBR bearers — GBR/MBR values.
Tying QoS to the bearer keeps the system simple: the eNB scheduler only ever looks at the bearer's QCI, not at the individual packets. Different applications with different needs get different bearers.
The bearer spans three concatenated segments — a radio bearer (UE–eNB), an S1 bearer (eNB–S-GW), and an S5/S8 bearer (S-GW–P-GW). A TFT (Traffic Flow Template) of packet filters decides which IP flows enter which bearer.
Because the bearer is the atom of QoS, the number of distinct treatments a UE can receive equals the number of bearers it holds. LTE allows up to 8 EPS bearers per UE (limited by the EPS Bearer Identity space usable for data), and every one of them is anchored on the default bearer of its PDN connection. When an application flow needs a treatment that no existing bearer provides, the network does not re-tag packets inside a bearer — it builds a new dedicated bearer whose QCI matches the requirement, then steers the flow into it with a TFT.
One line to remember: in LTE, one EPS bearer = one QCI. All packets in a bearer share the same treatment. 5G breaks this apart: many QFI flows can share one radio bearer, mapped by SDAP. LTE has no SDAP because there is nothing to map — the bearer is the QoS.
QCI — the standardised QoS label
The QCI (QoS Class Identifier) is a scalar that points to a set of standardised QoS characteristics. It is signalled as a single integer (values 0–255, with 1–9 and the extension values 65–85 standardised and the rest reserved or operator-specific). Because the values are standardised across the industry in TS 23.203, a node only needs to signal the QCI number and every other node already knows exactly how to treat that bearer — there is no need to send the characteristics explicitly across S1, S5/S8 or the radio interface. This is what makes QCI so compact and so powerful: a single integer replaces a whole descriptor of delay, loss and priority targets.
Each standardised QCI fixes four characteristics, plus (implicitly) a resource type:
| Characteristic | Meaning |
|---|---|
| Resource Type | One of GBR, Non-GBR, or (from Rel-13) Delay-critical GBR. It decides whether the bearer is granted a guaranteed bit rate reserved by admission control, or is best-effort and policed only by an aggregate cap. |
| Priority Level | Relative scheduling priority when the air interface is congested — lower value = higher priority. Standardised values span roughly 0.5 (mission-critical) to 9. The eNB scheduler uses it to arbitrate between bearers competing for the same resources. |
| Packet Delay Budget (PDB) | The upper bound (98th percentile) on the one-way delay between the UE and the P-GW. Standardised values range from 10 ms (delay-critical) to 300 ms (best-effort). It shapes how aggressively the scheduler and RLC must serve the bearer. |
| Packet Error Loss Rate (PELR) | The upper bound on the ratio of SDUs that are handed to the sender's link layer but not delivered — standardised values run from 10⁻² (voice) down to 10⁻⁶ (TCP data). It sets the residual reliability target, which drives HARQ and RLC (AM vs UM) configuration. |
Read direction of Priority Level: a classic trap. Lower number wins. QCI 5 at Priority 1 (IMS signalling) is served ahead of QCI 9 at Priority 9 (best-effort data). Think race position, not score.
The standardised QCI table
TS 23.203 Table 6.1.7 fixes a set of standardised QCIs so the whole network shares one meaning for each value. QCI 1–9 are the classic set that every commercial LTE network uses; later releases added 65/66/69/70 for mission-critical push-to-talk (MCPTT) and its signalling, 75/79 for V2X, and the 80/82/83 group for low-latency and ultra-reliable industrial services. Note that QCI 82/83 belong to the Delay-critical GBR resource type, where the PDB is interpreted more strictly (packets later than the budget count as lost against the PELR). The full standardised set:
QCI | Resource type | Priority | PDB | PELR | Example service |
|---|---|---|---|---|---|
| 1 | GBR | 2 | 100 ms | 10⁻² | Conversational voice (VoLTE) |
| 2 | GBR | 4 | 150 ms | 10⁻³ | Conversational video (live streaming) |
| 3 | GBR | 3 | 50 ms | 10⁻³ | Real-time gaming, V2X control |
| 4 | GBR | 5 | 300 ms | 10⁻⁶ | Non-conversational buffered video |
| 5 | Non-GBR | 1 | 100 ms | 10⁻⁶ | IMS signalling |
| 6 | Non-GBR | 6 | 300 ms | 10⁻⁶ | Buffered video, TCP data (premium) |
| 7 | Non-GBR | 7 | 100 ms | 10⁻³ | Voice, live video, interactive gaming |
| 8 | Non-GBR | 8 | 300 ms | 10⁻⁶ | Buffered video, TCP data (premium) |
| 9 | Non-GBR | 9 | 300 ms | 10⁻⁶ | Default bearer — best-effort internet |
| 65 | GBR | 0.7 | 75 ms | 10⁻² | Mission-critical push-to-talk voice (MCPTT) |
| 66 | GBR | 2 | 100 ms | 10⁻² | Non-mission-critical push-to-talk voice |
| 69 | Non-GBR | 0.5 | 60 ms | 10⁻⁶ | Mission-critical delay-sensitive signalling |
| 70 | Non-GBR | 5.5 | 200 ms | 10⁻⁶ | Mission-critical data |
| 75 | GBR | 2.5 | 50 ms | 10⁻² | V2X messages |
| 79 | Non-GBR | 6.5 | 50 ms | 10⁻² | V2X messages |
| 80 | Non-GBR | 6.8 | 10 ms | 10⁻⁶ | Low-latency eMBB (augmented reality) |
| 82 | Delay-critical GBR | 1.9 | 10 ms | 10⁻⁴ | Discrete automation (small payload) |
| 83 | Delay-critical GBR | 2.2 | 10 ms | 10⁻⁴ | Discrete automation (larger payload) |
Reading the table: VoLTE (QCI 1) tolerates a loose PELR of 10⁻² because one lost voice frame is inaudible, but needs a tight 100 ms PDB. Best-effort data (QCI 9) is the opposite — a generous 300 ms delay, but a strict 10⁻⁶ PELR because TCP hates loss. Mission-critical QCIs (65/69) get sub-integer priority values below 1 so they beat everything, and the 82/83 pair squeezes both delay (10 ms) and reliability (10⁻⁴) at once for factory automation.
ARP — admission and pre-emption
ARP (Allocation and Retention Priority) is a separate attribute of every EPS bearer, and it is deliberately not part of the QCI. Where the QCI describes how a bearer's packets are forwarded once the bearer exists, ARP governs whether the bearer gets set up at all when resources are scarce — and whether an existing bearer may be torn down to make room for a more important one. ARP is used at admission and retention time only; once a bearer is admitted, ARP has no further effect on day-to-day scheduling — that job belongs entirely to the QCI.
ARP has three fields:
| ARP field | Values | Meaning |
|---|---|---|
| Priority Level | 1–15 (1 highest) | Used at admission/retention to decide which bearer to admit or drop first under congestion. Values 1–8 are reserved for operator resources; 9–15 may be used for roamers/home subscribers. |
| Pre-emption Capability | enabled / disabled | May this bearer take resources already held by a lower-priority bearer when it is being set up? |
| Pre-emption Vulnerability | enabled / disabled | May this bearer's resources be taken away by a higher-priority bearer that needs to be admitted? |
QCI vs ARP: think of a restaurant. The QCI is how fast and carefully your meal is cooked once you are seated. ARP is the door policy — who gets a table when the place is full, and whose table can be given away. An emergency-call bearer carries a high ARP so it can pre-empt ordinary bearers (pre-emption capability enabled) and is itself protected (pre-emption vulnerability disabled).
Bit rates — GBR, MBR, and the AMBRs
GBR and non-GBR bearers are policed by different bit-rate parameters. GBR bearers reserve capacity end-to-end through admission control; non-GBR bearers reserve nothing and instead share an aggregate ceiling. Knowing which parameter applies to which bearer — and who enforces it — is a favourite interview topic.
| Parameter | Scope | Applies to | Meaning / enforcer |
|---|---|---|---|
GBR | Per GBR bearer | GBR only | Guaranteed Bit Rate — the rate the network commits to provide. Admission control reserves resources for it; enforced by the eNB scheduler. Signalled separately for uplink and downlink. |
MBR | Per GBR bearer | GBR only | Maximum Bit Rate — the ceiling for the bearer; traffic above MBR is discarded or delayed by the rate limiter (in LTE, MBR = GBR is common). Also signalled per direction. |
APN-AMBR | Per APN (per PDN connection) | Non-GBR only | Aggregate ceiling across all non-GBR bearers of one APN. Enforced in the uplink by the UE and in both directions by the P-GW. |
UE-AMBR | Per UE (across all APNs) | Non-GBR only | Aggregate ceiling across all non-GBR bearers of the UE; enforced by the eNB. It is capped at the sum of the subscribed APN-AMBRs the RAN will actually serve. |
Note: AMBR (Aggregate Maximum Bit Rate) applies only to non-GBR bearers. GBR bearers are governed by their own GBR/MBR and are not counted against the AMBR budgets. So a UE flooding a GBR VoLTE bearer never eats into the best-effort data budget, and vice versa.
The policy chain — PCRF, PCEF and PCC rules
The QoS values above do not appear by magic; they are decided by policy. The PCRF (Policy and Charging Rules Function) is the brain, and the PCEF (Policy and Charging Enforcement Function) — which lives inside the P-GW — is the muscle. TS 23.203 defines the split: the PCRF decides a policy for each service data flow and pushes it as a PCC rule; the PCEF in the P-GW enforces that rule by mapping the flow onto a bearer and gating/policing it.
The rule travels over the Gx reference point (Diameter) from PCRF to PCEF. A single PCC rule bundles: the SDF template (the 5-tuple packet filters identifying the service data flow — for example an IMS voice RTP stream), the QCI and ARP to apply, the GBR/MBR for GBR flows, and charging/gating information. When the PCEF receives a rule whose QCI/ARP does not match any existing bearer, it triggers a dedicated bearer activation: the P-GW sends a Create Bearer Request down through the S-GW and MME to the eNB and UE, carrying the bearer-level QoS and a TFT derived from the SDF filters. Once the bearer is up, the TFT installed at the P-GW (downlink) and the UE (uplink) steers matching packets into it.
The PCC rule is the unit of policy: SDF filters + QCI + ARP + GBR/MBR + gating/charging, decided by the PCRF.
Centralising policy in the PCRF lets the operator apply subscriber tier, application awareness and congestion state consistently, independent of the access node.
PCRF → PCEF over Gx; PCEF (in P-GW) installs the rule, activates a dedicated bearer, and enforces it with a TFT. The Rx interface lets an application function (e.g. the IMS P-CSCF) ask the PCRF for a rule.
Spec note: TS 23.203 places the PCEF in the P-GW and defines Gx (PCRF–PCEF) and Rx (AF–PCRF, e.g. IMS P-CSCF requesting QoS for a session). The bearer procedures the PCC rule triggers — Create/Update/Delete Bearer Request — are specified in TS 23.401.
From bearer to the radio — QCI to scheduling behaviour
Once a bearer is set up, its QCI characteristics drive everything below it. At the core/transport edge the P-GW marks packets (DSCP) and the S-GW forwards them per bearer; on the radio side each EPS bearer maps one-to-one to a DRB (Data Radio Bearer), which in turn maps to a MAC logical channel. The eNB scheduler and the MAC Logical Channel Prioritisation (LCP) procedure use the bearer's Priority Level, PDB, PELR and (for GBR) its GBR to decide how to serve it — GBR bearers are guaranteed their reserved rate first, then remaining capacity is shared among non-GBR bearers by priority, all while respecting each bearer's PDB.
This one-to-one bearer–DRB relationship is exactly what 5G relaxed: in NR, SDAP can multiplex several QoS flows onto one DRB. In LTE the scheduler's job is cleaner precisely because a logical channel is a QCI — it never has to inspect packets to know their treatment. See EPS Bearers and Bearer Management for how bearers are set up and modified.
LTE ↔ NR: the QoS attributes carry over but the granularity and packaging change. LTE's QCI becomes NR's 5QI (5G QoS Identifier), with the same four characteristics (resource type, priority, PDB, PELR) but a richer standardised table and support for both standardised and dynamically signalled 5QIs. LTE's ARP is kept almost unchanged. The big shift is structural: LTE ties QoS to the bearer (one QCI per bearer, one bearer per DRB), whereas NR ties QoS to a QoS flow (a QFI) and lets the new SDAP sublayer map many flows onto one DRB. So in NR the eNB/gNB scheduler can serve several QoS treatments over a single radio bearer — something LTE, with no SDAP, structurally cannot do.
Summary
LTE QoS is a per-bearer model with three moving parts. The QCI is a single standardised integer that stands for a whole treatment — resource type (GBR / non-GBR / delay-critical GBR), priority level (lower = higher), packet delay budget and packet error loss rate — so nodes exchange one number instead of a descriptor. The ARP is a separate per-bearer attribute (priority 1–15 plus pre-emption capability/vulnerability) that decides admission and pre-emption under congestion, not day-to-day scheduling. And bit rates split by resource type: GBR bearers get a reserved GBR and a ceiling MBR, while non-GBR bearers share the aggregate APN-AMBR and UE-AMBR caps.
Those values do not appear on their own: the PCRF decides a PCC rule and pushes it over Gx to the PCEF in the P-GW, which activates a bearer with the right QCI/ARP/GBR and installs a TFT. From there the QoS flows straight down — each EPS bearer is one DRB and one MAC logical channel, and the eNB scheduler plus LCP enforce the QCI characteristics every TTI.
Keep the anchoring principle in view: one bearer = one QCI. It is what makes the LTE scheduler simple (a logical channel is a QoS class) and it is precisely the constraint 5G lifts by moving to per-flow (5QI/QFI) QoS with SDAP.
Q. What is the finest granularity of QoS in LTE, and how is it identified?
A. The EPS bearer. It is identified by a single QCI plus an ARP; every packet in the bearer gets the same treatment. Contrast 5G, where the QoS flow (a QFI) is finer than the radio bearer.
Q. What does a QCI define, and what does it deliberately not include?
A. A QCI maps to Resource Type (GBR / non-GBR / delay-critical GBR), Priority Level, Packet Delay Budget and Packet Error Loss Rate. It does not include ARP — allocation/retention priority is a separate per-bearer attribute governing admission and pre-emption.
Q. Difference between APN-AMBR and UE-AMBR?
A. Both cap only non-GBR traffic. APN-AMBR limits the aggregate of non-GBR bearers within one APN (enforced by P-GW/UE); UE-AMBR limits the aggregate across all the UE’s non-GBR bearers and is enforced by the eNB.
Q. Who decides QoS policy and who enforces it, and over which interface?
A. The PCRF decides a PCC rule and pushes it over Gx to the PCEF inside the P-GW, which enforces it — activating a dedicated bearer with the rule's QCI/ARP/GBR and installing a TFT to filter the service data flow into it.
Q. How does a packet end up on a dedicated bearer rather than the default one?
A. A TFT of packet filters (a 5-tuple SDF template from the PCC rule) matches the packet; matching packets ride the dedicated bearer, and anything that matches no dedicated TFT falls through to the default bearer (typically QCI 9).
Q. How does LTE's QCI relate to 5G's QoS identifiers?
A. The QCI becomes the 5QI in NR — same four characteristics, but with a larger standardised table and support for dynamically signalled values. The structural change is that NR attaches QoS to a per-flow QFI and maps flows to DRBs via SDAP, instead of LTE's one-QCI-per-bearer model.
Where this connects
QoS in LTE lives on bearers, so the natural next steps are how those bearers are built, changed, and used by a real service: