ICCRG - Congestion control for 4G/5G networks

Congestion control for 4G/5G networks

Ericsson Research - Ingemar Johansson

ICCRG - Congestion control for 4G/5G networks | 2016-07-01 | Page 2

› Don’t rely too much on RTT measurements

› Inter-arrival estimates may be noisy

› Packet pacing improves stability

› Delay based CC’s can be good fallback to combat bufferbloat

› High peak throughput or low latency (under load) ?, pick one

› Warm up your device before ping measurements

Congestion control for 4G/5GSome considerations


Segmentation, ARQ

Ciphering

Header Compr.

Hybrid ARQHybrid ARQ

MAC multiplexing

Antenna and resrouce mapping

Coding + RM

Data modulation

Antenna and resource mapping

Coding

Modulation

Antenna and resource

assignment

Modulation

scheme

MA

C s

ch

ed

ule

r

Retransmission control

Priority handling, payload

selection

Payload selection

RLC

PHY

PDCP

User #i User #j

Reassembly, ARQ

Deciphering

Header Compr.

Hybrid ARQHybrid ARQ

MAC demultiplexing

Antenna and resrouce mapping

Coding + RM

Data modulation

Antenna and resource

demapping

Decoding

Demodulation

RLC

PHY

PDCP

MAC

Red

un

da

ncy

ve

rsio

n

Radio Bearers

Logical Channels

Transport Channel

MAC

MAC

The LTE Protocol stack


› Parallel stop-and-wait processes

– 8 processes 8 ms roundtrip time

› ~10% probability of retransmission on MAC layer

– Gives 8ms extra latency

– RLC implements in sequence delivery

MAC LaYER HARQHybrid-ARQ with Soft Combining

Process transport block 3 Process transport block 5

Process #0

Process #1

Process #2

12 3 5 14

Process transport block 2 Process transport block 4

Process transport block 1 Process transport block 1 Process transport block 1

Hybrid-ARQ

protocol

To RLC for in-sequence delivery

Block 2

Process #7

Block 3 Block 4 Block 5 Block 1

1


› Scheduler behavior is not standardized, implementations are vendor

specific

› Different kinds of schedulers

–Delay : VoLTE (voice)

–Proportional fair, Round Robin : Default (=best effort) bearers

–Max CQI..

› Scheduling has impact on IP packet delay characteristics

DL and UL schedulinG


› Downlink queuing delay show

little tendency to increase

› Uplink delay is however relatively

large, why ?

Example 1

0

0.5

1

1.5

2x 10

5 CWND

0 0.05 0.1 0.15 0.20

0.05

0.1

T[s]

IP packet delay [s]

Downlink

Uplink


› Problem free packet

transmission in downlink

› TCP bursts are transmitted as

quickly as resource allocation

allows

Example 1, Sequence numbers DL

0

1

2

3x 10

5 TCP sequence numbers

0

100

PDCP sequence numbers (DL)

0 0.05 0.1 0.15 0.20

50

100

T[s]

RLC sequence numbers (DL)

TX DL

RX DL


› Uplink ACKs are bundled RTT

affected by UL scheduling

› An uncongested downlink may

thus be considered as congested

by a too sensitive RTT based

congestion trigger

– Example : Cubic HyStart

Example 1, Sequence numbers UL

0

1

2x 10

5 TCP ACKed sequence numbers

0

50

PDCP sequence numbers (UL)

0 0.05 0.1 0.15 0.20

10

T[s]

RLC sequence numbers (UL)

TX UL

RX UL

SC Req

SC Grant

Don’t rely too much on RTT estimates


› Gaps in IP packet delay, why ?

Example 2

0

5

10

x 104 CWND

0 0.05 0.1 0.15 0.20

0.02

0.04

0.06

0.08

T[s]

IP packet delay [s]


› RLC AM bearer (in-sequence

delivery)

› HARQ retransmissions delay

packet forwarding to higher layers

Example 2, Sequence numbers DL

0

1

2x 10

5 TCP sequence numbers

0

50

100

150PDCP sequence numbers (DL)

0 0.05 0.1 0.15 0.20

50

100

T[s]

RLC sequence numbers (DL)

TX DL

RX DL

Inter packet arrival measurements can become noisy


› LTE transport block sizes are limited

– Example : 10MHz BW, max TB size ~36kByte

› Large bursts are queued up, queuing delay can trigger unnecessary AQM drops

or ECN marks

› Lower AQM/ECN thresholds Higher risk of false congestion indication

– Consider L4S

Packet pacing


Bursty traffic example

› Video streaming example

› Bursty on/off pattern generates

large amounts of immediate

data in RLC queue

› Leads to queuing delay spikes

when new burst starts AQM

packet discard Video rate

reduction

› Solution : Enable packet pacing

44 45 46 47 48 49 50 51 520

0.1

0.2

0.3

0.4Queuing delay [s]

44 45 46 47 48 49 50 51 520

200

400

600

800

1000Congestion window [MSS]

T [s]

New burst,

600packets ~880kB

gives delay spike

AQM discard

CWND is reduced


Smaller risk of unnecessary AQM discards or ECN marks

with packet pacing

Packet pacing example

0

2

4

6

8x 10

4 CWND

0 0.05 0.1 0.15 0.20

0.05

0.1

T[s]

IP packet delay [s]

Downlink

Uplink

0

2

4

6

8x 10

4 CWND

0 0.05 0.1 0.15 0.20

0.05

0.1

T[s]

IP packet delay [s]

Downlink

Uplink

Packet pacing OFF Packet pacing ON

More consistent

queue buildup


› Throughput variations

– Larger variations expected

› Carrier Aggregation

› Dual Connectivity

› Multi-beam technology

– Solutions :

› Fast increase options

- SIAD

- Cubic with fast increase option

› Other : PCC, Verus ?

› Bufferbloat

– Expected to be solved over time

› AQM, ECN, L4S

– But fallback to a delay/loss based CC good

Congestion control for 4G/5G

../../Presentations/External presentations/mobilityMovie.avi


5 10 15 20 25 300

0.5

1

TCP segment delay [s]

5 10 15 20 25 300

1000

2000

Congestion window [kbyte]

5 10 15 20 25 300

20

40

Throughput [Mbps]

T [s]

5 10 15 20 25 300

0.5

1


5 10 15 20 25 300

1000

2000


5 10 15 20 25 300

20

40

Throughput [Mbps]

T [s]

5 10 15 20 25 300

0.5

1


5 10 15 20 25 300

1000

2000


5 10 15 20 25 300

20

40

Throughput [Mbps]

T [s]

Example : Cubic modificationRTTmin = 100ms , Variable BW = 5505Mbps

Cubic +fast increase +delay target 100ms

Modified congestion controls can improve performance in

typical 4G/5G scenarios


Peak throughput vs low delay10MB file transfer

0

0.05

0.1


0

500Congestion window [kB]

0

10

20

Throughput [Mbps]

0 5 10 15 20 25 30 350

500

1000RLC queue [bytes]

T [s]

0

0.05

0.1


0

100

200

Congestion window [kB]

0

10

20

Throughput [Mbps]

0 5 10 15 20 25 30 350

500

1000RLC queue [bytes]

T [s]

ECN th. 100ms ECN th. 20ms

A too shallow

RLC queue

gives degraded

resource utilization

A low target queue delay can give poor link utilization

Reason : Too little data in RLC queue


Latency in 5G ?

› Numerology

– Short subframe duration, 0.25-0.5 ms or less low latency (together with tighter UE/eNB processing)

› HARQ RTT reduction from ~8ms to ~1ms possible

– More robust HARQ

– Multiple numerologies may be needed

› Depending on deployment scenario, use case and frequency band

– Many alternatives

› Option for contention based scheduling in UL

15 kHz20 MHz 1 ms

Phase-noise robustness,

reduced CP overhead

Reduced latency

Increased bandwidth Increased subcarrier spacing Reduced subframe duration

Very high data rates


Questions/comments?

65.359466N, 22.478387E

Extra slides

More on latency in LTE


› Two ECM states

– ECM-IDLE: No physical resources assigned

– ECN-CONNECTED: Physical resources assigned

› UE goes into ECM-IDLE if inactive for a longer time

– (vendor specific, 10s, 60s...)

LTE states

Physical resource : Radio resource (SRB/DRB) and

Network Resource (S1 bearer/S1 signalling connection)


› C-DRX = Connected Mode DRX (ECM-CONNECTED)

› UE radio is turned off if nothing transmitted or received for an inactive timer

period (e.g 10ms)

– Radio turned on immediately if UL data available in UE

› UE wakes up and listens at intervals given by a short DRX timer (e.g 40ms)

› If no data transmitted or received for a given number of short DRX cycles, go

into long DRX

– UE wakes up and listens at intervals given by a long DRX timer (e.g 320ms)

› If UE inactive for a longer time (10s-60s) go into idle state (ECM-IDLE)

– UE wakes up and listens for paging at standardized intervals (e.g 1024ms)

DRX Discontinuous Reception


› LTE– Connection establishment

– Synchronization

– Air transport

– Processing

– Paging/C-DRX (DL)

– Handover/Mobility

› EPC– Connection establishment

– Handover/Mobility

– Packet processing (very small)

› Transport network– Propagation delay

– Processing

› Application– Processing delay

› Internet transport– Propagation delay

– Processing

LTE Latency breakdown

EPC eNB App UE

Conditional:

Nothing/Synchronization/

Connection establishment

Trigger

Air transport +

processing

Propagation +

processing Internet

Propagation +

processing

Application

processing

Application

processing


LTE UL air interface latenciesexample 1

SR

Grant

Data + BSR

Data

Data is created

and packetized

Data

Data

Grant

Data

Data

Grant

Data

Data

Data

T1

T2

T3a

T3b

T4

T5

T6

T7

T3

T0

Modem eNB SGWMME PGWApp

Server

UE L2 Processing

+ SR periodicity

SR transmission

+ eNB processing

Grant

transmission +

UE L2

processingData (incl BSR

reporting) + eNB

processing

eNB

processing

and grants

(conditional

due to BSR)

Data (incl BSR

reporting) + UE

and eNB

processing

T0

T1 0.5-5 ms

T2 1 ms

T3a 2 ms

T3b 0 ms

T4 1 ms

T5 3 ms

T6 1 ms

T7 2 ms

Total: 10.5-15 ms

From ECM-CONNECTED


LTE UL air interface latenciesexample 2

UE in ECM-IDLE

Data

Initial UE Message

Data packet

Modify Bearer Resp

Modify Bearer Req

Modify Bearer Resp

Modify Bearer Req

T1RRC Connection

Request

RRC Connection

Set-up

RRC Connection

Set-up Complete

Security Mode Comm.Initial Context

Setup Req.Security Mode Comp.

Initial Context

Setup Resp.

UE in ECM-CONNECTED

Conditional: UE Capability Retrieval (not included)

RRC Connection

Reconf.RRC Connection

Reconf. Complete

Optional: Authentication and NAS Security Activation (not included)

RACH preamble

RACH Response

Modem eNB SGWMME PGW HSS

T2

T3

T4

Data

T0

T5

From ECM-IDLE: 130 ms – Don’t mix states in testing


LTE DL AIR Interface latenciesexample 1

T1a 2 ms

T1b 0 ms

T1c 0.5 ms

T2 1 ms

T3

Total: 3.5 msData packet

to higher

layers

T1a

T2

Data

Data packet

Data

T1b

Data

T3

T1

T1c

Modem eNB SGWMME PGW

eNB processing

including frame

alignment

scheduling and C-

DRX cycle

Data transport

From ECM-CONNECTED


Data packet

to higher

layers

Preamble on

PDCCH

Data packet

Data

Data

RACH

Preamble

RA

Response

Data

T1a

T1b

T2

T3

T4

T5

T6

T7

T8

T9

Modem eNB SGWMME PGW

From ECM-CONNECTED but out of synch: 23 ms – Don’t mix states in testing

LTE DL AIR Interface latenciesexample 2


0

Pre-scheduling Pre-scheduling OFF

Delay Breakdown

Time [ms]

SR

Grant

PING

PINGPING

PING

PINGPING

5 SR encod. and alignment

1

eNB processing3

1

11

11

3GPP specified delay4

Core network delay3

2

23 ms Total round trip delay *

5

10

15

20

*Note: Numbers are example only


10

0

Pre-schedulingPre-scheduling ON

Delay Breakdown

Time [ms]

GrantPING

PINGPING

PING

PINGPING

11

11

3GPP specified delay4

Core network delay3

2

13 ms Total round trip delay *

5

15

20

*Note: Numbers are example only


› The gain of the prescheduling on short file

transfers is clear and directly proportional to

the period of the prescheduling.

› The delta between best and worst case

scenario gains is around 100 ms

› Cost : More interference and resource usage

Pre-scheduling Short file download

No Prescheduling

2 ms period

No Pre-scheduling

2ms

period

5ms

10ms20ms


› EPS : Evolved Packet System

› EPC : Evolved Packet Core

› EMM : EPS Mobility Management

› ESM : EPS Session Management

› ECM : EPC Connection Management

› SRB : Serving Radio Bearer

› DRB : Data Radio Bearer

› DRX : Discontinuous Reception

› UE : User Equipment ~ Terminal

Glossary

Documents

ICCRG - Congestion control for 4G/5G networks