A Machine Learning Based Approach to Mobile Network...

A Machine Learning Based Approach to Mobile Network Analysis

Zengwen Yuan1, Yuanjie Li1, Chunyi Peng2, Songwu Lu1, Haotian Deng2, Zhaowei Tan1, Taqi Raza1

1 2

Overview�2

Background

Overview�2

Background

Why machine learning for mobile network analysis

Overview�2

Background


Mobile network analysis: state-of-the-a; and our approach

Overview�2

Background



Case study: analyzing latency for mobile networks

• How mobile apps work over LTE

• How to breakdown app-perceived latency

• Challenges and ML scheme

• Primary results from crowdsourcing

Overview�2

Background



Case study: analyzing latency for mobile networks

• How mobile apps work over LTE

• How to breakdown app-perceived latency

• Challenges and ML scheme

• Primary results from crowdsourcing

Conclusion

Overview�2

Ubiquitous cellular networks connect everyone, everything

�3

The race to 5G opens many new oppoCunities�4

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application StackWeb

server

Internet??

Cellular network (4G LTE)User space

Yet, access to mobile network analytics is barred�5

What’s going on in the 3G/4G/5G network??

Oh we cannot tell you unless you sign an NDA…

Researcher (you)

No privilege no talk, sorry!

Chipset/Mobile OS

Operators

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface


server

Internet??

Cellular network (4G LTE)User space

Yet, access to mobile network analytics is barred�5

What’s going on in the 3G/4G/5G network??

Oh we cannot tell you unless you sign an NDA…

Researcher (you)

No privilege no talk, sorry!

Chipset/Mobile OS

Operators

3G/4G operations remain closed both in device chipsets and network infrastructures :(

Plus, mobile networks are complex & distributed

More complex functions on both control and data planes

Operations are distributed across layers

�6

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

TCP/IPProtocol Stack

Application Protocols

Radio Resource Control (RRC)

Mobility Management (EMM)

Session Management (ESM)

Packet Convergence (PDCP)

Radio Link Control (RLC)

Medium Access Control (MAC)

Physical Layer Protocols (PHY)

Control Plane

Data Plane

Signaling

Moreover, analytics tasks are app speciNc�7

Analytics for mobile networks is problem-speciCc, for example:



• Web browsers:✦ Why the time-to-first-byte (TTFB) is so long?

✦ What’s the major component of latency?

✦ …



• Web browsers:✦ Why the time-to-first-byte (TTFB) is so long?

✦ What’s the major component of latency?

✦ …

• Instant message apps:✦ Does the recipient read my message?

✦ Is my message delivered in time?

✦ …


State-of-the-aC mobile network analytics�8

Current 4G network analytics is primarily “infrastructure-based”:




Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface


server

Internet

Cellular network (4G LTE)

Base stations

Gateway

Gateway

User profile server

Mobility controller

User space



Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface


server

Internet


Base stations

Gateway

Gateway

User profile server

Mobility controller

User space

Not Scalable



Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface


server

Internet


Base stations

Gateway

Gateway

User profile server

Mobility controller

User space

Not Scalable Incomplete View



Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface


server

Internet


Base stations

Gateway

Gateway

User profile server

Mobility controller

User space

Not Scalable Incomplete View Opaqueness

ML-based approach is a must-have feature for mobile network analytics

�9


Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Device-centric ML approach brings new hope


�9


Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space



�9


Scalability

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space



�9


Scalability Device QoE View

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space



�9


Scalability Device QoE View Availability

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

It is probably true that machine learning is a must-have approach, rather than a nice-to-have one, to our Celd for mobile network analysis

Our proposal: two-level device-centric ML approach �11

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Local level: sensing mobile network data inside each sma;phone

• Via hardware-software coordination (e.g. MobileInsight [ACM MobiCom’16])

• Via higher-layer (application/transport/IP) and lower-layer (cellular-specific) integration

• Via ML-assisted data plane prediction from control plane protocol reconstruction


Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Local level: sensing mobile network data inside each sma;phone

• Via hardware-software coordination (e.g. MobileInsight [ACM MobiCom’16])

• Via higher-layer (application/transport/IP) and lower-layer (cellular-specific) integration

• Via ML-assisted data plane prediction from control plane protocol reconstruction

Global level:

• Crowdsourcing-based dataset

• Cloud-synthesized insights


Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Mob

ile A

pps

Ba

seba

nd

Mob

ile O

S

TCP/

IP s

tack

LTE

inte

rfac

e

Appl

icat

ion

Stac

k

User

spa

ce

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Step 1: open up the “black-box” operations

• At/above IP network data: TCPDUMP

• Below IP network data: MobileInsight

Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space




Step 2: automated data preprocessing

• Data cleansing and integration of two sources

Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space






Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Preprocessing






Step 3: local ML-based analysis

• Control plane for protocol operations

• Data plane for performance

Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Preprocessing






Step 3: local ML-based analysis

• Control plane for protocol operations

• Data plane for performance

Local analysis�12

Mobile Apps

Baseband

Mobile OS

TCP/IP stack

LTE interface

Application Stack

User space

Preprocessing ML analysis

Global analysis

Enabled by cloud-based crowdsourcing (e.g. cniCloud [HotWireless’17])

Analytical Insights for:

• Geographical location

• Operators

• Phone models

• …

�13

…

SQL Query

SQL Response

Fine-grainedlogging&sharing

EfficientDataManagement

StructuredQuery

Case study: latency analysis in mobile networks

Every millisecond of latency maSers!

Mobile network users want fast access

• 1 second latency in page response → 7% reduction in PageView [KissMetrics 2011]

Developers lose revenue due to long latency

• Every 100 ms costs Amazon 1% ($1.6 bn) in sales

• An extra 400 ms latency drops daily Google searches per user by 0.6%

Latency does maZer a lot!

�15

Background: how do mobile apps work over 4G LTE?

What happens under the hood?

How LTE impacts perceived latency on mobile web/IM app?

�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface

Application(HTTP/DNS)

Web server

Internet




�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet?




�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet





�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

Base stations

Gateway

Gateway

User profile server

Mobility controller





�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

(a)

(b)

Base stations

Gateway

Gateway

User profile server

Mobility controller





�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

(a)

(b) (c)

(c)

(c)

Base stations

Gateway

Gateway

User profile server

Mobility controller


(d)

(d) (d)




�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

(a)

(b) (c)

(c)

(c)

Base stations

Gateway

Gateway

User profile server

Mobility controller


(d)

(d) (d)




�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

(a)

(b) (c)

(c)

(c)

Base stations

Gateway

Gateway

User profile server

Mobility controller

(f)

(e)


(d)

(d) (d)




�16

Safari WhatsApp

modem chipset

mobile OS

TCP/IP stack

LTE interface


Web server

Internet

(a)

(b) (c)

(c)

(c)

Base stations

Gateway

Gateway

User profile server

Mobility controller

(f)

(e)

LTE control-plane operations pose sizable latency on mobile apps

Timing breakdown of control plane operations�17


P1a: RRC connection setup request



(Random Access)



(Random Access)

P1b: RRC connection setup



(Random Access)


P1c: RRC connection setup complete



(Random Access)



P2. Service request


P3. Authentication (non-mandatory)


(Random Access)



P2. Service request



P4. Initial Context Setup


(Random Access)



P2. Service request


P5a. Security mode command




(Random Access)



P2. Service request






(Random Access)



P2. Service request

P5b. Security mode complete






(Random Access)



P2. Service request


P6a. RRC connection reconfig






(Random Access)



P2. Service request



P6b RRC connection reconfig complete






(Random Access)



P2. Service request




Data bearer Data bearer






(Random Access)



P2. Service request




Data bearer Data bearerUL data






(Random Access)


P1c: RRC connection setup completeT

P2. Service request










(Random Access)


P1c: RRC connection setup completeT

P2. Service request




T


Learning latency: latency data sensing�18

Three-tiered timing data collection:

• App-specific semantic timing (e.g. Navigation Timing API, IM timing model)

• TCP/IP stack timing (from TCPDUMP)

• LTE stack timing (from MobileInsight)

DNS query TCP connection

TCP SYN

Data access request

LTE data

SYN ACK

HTTP request

LTE control plane

LTE data plane

OS overhead

HTTP transmission

Pagerendering

TCP dataTCP layer

unloadEventStartfetchStart

domainLookupStartdomainLookupEndconnectStartconnectEnd

requestStart

responseEndresponseStart

domInteractive

loadEventEnd

Challenge: timestamp alignment�19

How to align timestamps at these layers?

• Domain-specific event tracing and mapping

• Machine-learning assisted

App TCP/IP LTE chipset

Base station

Web server

LTE control plane msgs

network request

LTE data plane pkts Server’s ACK

TappTtcpdump

TLTE

App log tcpdump

log MobileInsightlog

network response

Pinpoint latency boSleneck in LTE: An example�20

Pinpoint latency boSleneck in LTE: An example

Run a small webpage (4 KB) in Chrome on Android

• User is static, under good 4G LTE signal (-95 dBm), T-Mobile

�20




Total Latency: 473 msec

• Clicking URL → page loading complete, Steps (a)—(f)

�20




Total Latency: 473 msec

• Clicking URL → page loading complete, Steps (a)—(f)

Pinpointing the latency boZleneck

• How to breakdown?

�20

Control-plane latency breakdown: local analysis I�21

Control-plane latency breakdown: local analysis I

Major component from Navigation Timing API: DNS lookup, 250 ms out of 473 ms�21


Major component from Navigation Timing API: DNS lookup, 250 ms out of 473 ms

Is the DNS server slow to handle connection?

�21




Fu;her breakdown: LTE service request takes 172 ms before the DNS setup

�21




Fu;her breakdown: LTE service request takes 172 ms before the DNS setup

�21

QueueingStalledDNS LookupInitial ConnectionRequest SentWaiting (TTFB)Content Download

8.98 ms2.97 ms250.04 ms30.11 ms0.36 ms137.41 ms43.48 ms

LTE Data Access LatencyLTE service request

Data-plane latency breakdown: local analysis II

Fu;her zoom in and breakdown the remaining LTE data access latency (291 ms):�22

DNS-Wait GrantDNS (IPv6)DNS-Wait GrantDNS (IPv4)APP-OS overheadTCP SYN-Wait GrantTCP SYN-Send DataTCP ACK (local processing)HTTP GET (send request)HTTP GET-Wait GrantHTTP GET-req sentHTTP-server RTT+ DL latencyLTE-to-TCP overheadHTTP page DL transmissionHTTP DL retransmission

2.02 ms11 ms

18 ms0.02 ms

First bit of HTTP response

0.36 ms12 ms

8 ms110 ms6.1 ms

40 ms3 ms

16 ms

17 ms26 ms

12 ms

Latency mapping for failures: local analysis III

Example: data plane suspension due to radio reconnection and head-of-line blocking during handover

�23

Latency mapping for failures: local analysis III

Example: data plane suspension due to radio reconnection and head-of-line blocking during handover

�23

BlockingRequest SentWaiting GrantUplink TransmissionHandover Disruption — No dataHandover Disruption — Duplicate recv’d dataWaiting (TTFB, due to parallel TCP connection)Content Download

5.05 ms0.58 ms4 ms130 ms263 ms36 ms275 ms33.16 ms

Machine learning scheme

We leverage domain-speciCc knowledge for ML-based predictions

Control plane: predict handover using a decision tree classiCer

• Features from 3GPP standards

• Predicts handover 100ms before it occurs with >99% accuracy

Data plane: predict NACK/ACK hip at MAC layer

�24

Synthesizer: global crowdsourcing analysis�25

Synthesizer: global crowdsourcing analysis

Four US carriers + Google Project Fi�25


Four US carriers + Google Project Fi

23 phone models, 95,057 data sessions

�25




Overall latency: 77 — 2956 ms in 500K samples

�25





• Varies among different mobile carriers

�25

Average Latency by LTE Data Access Setup (no mobility)

0

50

100

150

200

AT&T T-mobile Sprint VerizonProject Fi

162153147165

196





• Varies among different mobile carriers

• Insensitive to varying radio link quality

�25

-130-120-110-100 -90 -80 -70 -60 -50 -40

50

100

200

500

1,000

3,000

Signal Strength (dBm)

TotalLatency(ms)

LTE data access latency: how frequent?

Frequent data access setup operations

• every 58.8 sec (median); 133.6 sec (average)

• cause: frequently entering power-saving mode

Sho1-lived Radio connectivity lifetime

• every 10.8 sec (median); 17.3 sec (average)

• cause: inactivity timer (regulated by standards)

�26

Overall latency and breakdown for major carriers�27

AT&TT-Mobile SprintVerizon Project Fi

Findings Summary

Tradio: Radio connectivity setup

• It contributes 67.5 −1665.0 ms of the overall LTE access latency.

• On average, it contributes 39.7%, 44.0%, 61.9%, 64.2% and 43.7% of total latency in T-Mobile, AT&T, Verizon, Sprint and Project-Fi, respectively.

Tctrl: Connectivity state transfer

• It contributes 28.75 ms to 2286.25ms of the overall LTE access latency.

• On average, it contributes 60.3%, 56.0%, 38.1%, 35.8% and 56.3% of total latency in T-Mobile, AT&T, Verizon, Sprint and Project-Fi, respectively.

�28

Impact on mobile Web app: Chrome

Average page loading time for tested webpage: 411 ms

• LTE data access setup: 174 ms

• 42.3% total latency perceived

Similar results for Safari latency on iOS

�29


TCP SYN

Data access request

LTE data

SYN ACK

HTTP request

LTE control plane

LTE data plane

OS overhead

HTTP transmission

Pagerendering

TCP dataTCP layer

unloadEventStartfetchStart

domainLookupStartdomainLookupEndconnectStartconnectEnd

requestStart

responseEndresponseStart

domInteractive

loadEventEnd

0 25 50 75 1000

200

400

600

800

Normalized sorted sample (%)

Latency(ms)

Total latency

LTE latency

Impact on instant-messaging: WhatsApp

Average time Crst data packet being ACKed: 341 ms

• LTE data access setup: 175 ms

• 51.4% total latency perceived

�30


SYN

Data access request

LTE data

SYN ACK

LTE control plane

LTE data plane

OS overhead SSL Data

TCP dataTCP layer

App init

TCP ACK

SSL Data

TCP ACK

App connect w/ server

New messageidle

Server ACK

Latency

Next message

0 25 50 75 1000

200

400

600

800

Normalized sorted sample (%)

Latency(ms)

Total latency

LTE latency

Discussion: reducing LTE latency

Data plane walk-arounds

• Mask the data setup latency by waking device in connected mode in advance

Control plane acceleration

• Speed up connectivity state transfer between the base station and the mobility controller (e.g. DPCM [ACM MobiCom’17])

• Handover prediction

Other issues

• Extending to other network metrics (e.g. loss, throughput, …)

• Theoretical bounds

• Privacy issues

�31

Conclusion: ML-based analysis for next-gen mobile networks

Mobile networks are successful and will continue to prosper (5G, self driving, …)

Mobile network analysis: paradigm shin to device-centric, ML-based scheme

• Device-centric: unveil the tightly-guided operation issues over 4G/5G mobile networks

• Two-tiered approach: a more open solution approach for the research community

�32

Backup

A Machine Learning Based Approach to Mobile Network...

Documents

Transcript of A Machine Learning Based Approach to Mobile Network...