Toward Next-Gen Low Latency Mobile Networks Zengwen Yuan - - PowerPoint PPT Presentation

Toward Next-Gen Low Latency Mobile Networks

Zengwen Yuan University of California, Los Angeles 2018-05-06

WiNG

Wireless Networking Group

More time spent on mobile Internet than ever

2

Every millisecond counts in delivering mobile data

Users want fast mobile network experience

• 1 second latency in page response → 7% reduction

in PV [KissMetrics 2011]

Websites 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 ma:ers!

3

Latency peAormance: are we there?

If you ever encountered slow loading webpage… Common sense: radio quality/data bandwidth dominates latency Why perceivable latency still observed when radio quality/bandwidth is good in 4G LTE networks?

• What is the contributing latency bottleneck?
• How to reduce latency with new design?

4

Not aways true!

Outline

What is the contributing latency bo:leneck?

• LTE control-plane latency: An often overlooked bottleneck

How to reduce latency with new design?

• DPCM: A control-plane perspective in latency reduction [ACM MobiCom’17]

5

Cellular network (4G LTE)

(d) (d) (d)

Background: how do mobile apps work over LTE?

What happens under the hood? How LTE impacts perceived latency on mobile app?

6

Safari WhatsApp

modem chipset mobile OS

TCP/IP stack

LTE interface

Application (HTTP/DNS) Web server

Internet

(a) (b) (c) (c) (c)

Base stations Gateway Gateway User profile server Mobility controller

?

(f) (e)

LTE control-plane operations precede data forwarding

A Close Look on Latency BoIleneck

Pinpoint latency boIleneck 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

Total Latency: 420 msec

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

Pinpointing the latency bo:leneck

• How to breakdown?

8

DNS query

TCP connection

HTTP request

OS

• verhead

HTTP transmission Page rendering

Latency model for Web using HTML APIs

Latency component breakdown

Major component: TCP connection setup, 283 ms out of 420 ms (67 %) Is the server slow to handle connection?

• No. It is 3x longer than average phone-server roundtrips (HTTP RTT, 71ms)

Through fu\her breakdown, LTE control-plane procedures take 172 ms (40% of total latency), during TCP connection setup

9

DNS query

TCP connection TCP SYN

Data access request

LTE data

SYN ACK HTTP request

LTE control plane LTE data plane

OS

• verhead

HTTP transmission Page rendering

TCP data

TCP layer

LTE control-plane operations incur sizable latency for mobile Web and other apps

An oLen overlooked latency source: LTE C-plane

A new source of latency — LTE control plane operations Our 20-month user study [1] unveils:

• 148 ms to 196 ms average data access setup latency among five US carriers
• It may go up to 2.96 seconds in rare cases

Visible impact on apps:

• 40.4% of overall web latency on average
• 51.4% of overall instant-messaging app latency on average

10

[1] 5M 4G signaling messages collected by MobileInsight dataset at http://mobileinsight.net

LTE data access latency: how long?

77 — 2956 msec in 500K samples

• Varies among different mobile carriers
• Insensitive to varying radio link quality

11

Average Latency by LTE Data Access Setup (no mobility)

50 100 150 200 AT&T T-mobile Sprint Verizon Project Fi

162 153

147

165 196

LTE data access latency: how long?

77 — 2956 msec in 500K samples

• Varies among different mobile carriers
• Insensitive to varying radio link quality

12

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

50 100 200 500 1,000 3,000

Signal Strength (dBm) Total Latency (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

Sho4-lived Radio connectivity lifetime

• every 10.8 sec (median); 17.3 sec (average)
• cause: inactivity timer (regulated by standards)

13

Impact on mobile Web app: Chrome

Average page loading time for tested webpage: 319 ms

• LTE data access setup: 129 ms
• 40.4% total latency perceived

Similar results for Safari latency on iOS

14

DNS query

TCP connection TCP SYN

Data access request

LTE data

SYN ACK HTTP request

LTE control plane LTE data plane

OS

• verhead

HTTP transmission Page rendering

TCP data

TCP layer

unloadEventStart fetchStart domainLookupStart domainLookupEnd connectStart connectEnd requestStart responseEnd responseStart domInteractive loadEventEnd

25 50 75 100 200 400 600 800

Normalized sorted sample (%) Latency (ms)

Total latency LTE latency

Impact on instant-messaging: WhatsApp

Average time grst data packet being ACKed: 341 ms

• LTE data access setup: 175 ms
• 51.4% total latency perceived

15

DNS query

TCP connection

SYN

Data access request

LTE data SYN ACK

LTE control plane LTE data plane

OS

• verhead

SSL Data TCP data

TCP layer

App init

TCP ACK

• SSL

Data TCP ACK

• App connect

w/ server New message idle Server ACK

Latency

Next message

25 50 75 100 200 400 600 800

Normalized sorted sample (%) Latency (ms)

Total latency LTE latency

DPCM: A New Design for Low Latency Mobile Network*

*ACM MobiCom 2017, Joint work with Yuanjie Li and Chunyi Peng http://metro.cs.ucla.edu/dpcm.html

Sequential LTE control-plane is slow

Example: LTE data access setup as the device moves

17

Data-plane Control-plane

Mobile Network (4G LTE)

New Location Old Location

Data Access Setup Latency

• P4. Routing path update
• P5. User profile location update
• P3. Authentication & security
• P2. Session state migration (QoS, billing, security, …)
• P1. Radio connection setup

UL/DL data service available Non-mandatory procedures block data Parallelizable procedures are forced to run sequentially Failures block the entire

• perations

Sequential, slow!

Accelerating LTE control-plane

Challenges

• 1. How to retain the control/data-plane correctness?
• 2. How to efficiently perform acceleration?

18

From a distributed state-mgmt point of view

19

Location IP QoS Security Billing

ACL

Security IP QoS Security Billing

ACL

IP QoS Billing QoS Security Billing

ACL

Location QoS Security Billing

ACL

IP QoS Billing Location IP QoS Security Billing

ACL

Location Location

Device-side state replica: Almost “always-on” Network-side state replica: On multiple nodes

Accelerating LTE control-plane (cont’d)

Challenges

• 1. How to retain the control/data-plane correctness?
• 2. How to efficiently perform acceleration?

Insights

• 1. Accelerate via distributed state management

✦ Correctness achieved via equivalent but faster state operations

• 2. Use state replicas that readily exist on device and in network

✦ Device and network nodes have up-to-date state replicas in many cases ✦ Without waiting for a single node (mobility controller) to become bottleneck

20

DPCM: a new perspective on data latency in LTE

We revisit LTE control-plane functions from distributed state mgmt perspective

21

Non-mandatory procedures block data Parallelizable procedures are forced to run sequentially Failures block the entire

• perations

Pipeline control procedures w/ data forwarding Parallelize multiple control procedures Bypass slow bottleneck control procedures

Location QoS Security Billing ACL IP QoS Billing Location

Acceleration with State Replicas:

Bypass the Slow Control Procedures

Data-plane Control-plane

Location Location IP QoS Security Billing

ACL

Security QoS Security Billing

ACL

IP QoS Security Billing

ACL

IP QoS Security Billing

ACL

Location Location

Shorter side-path with device-side state replica

Pipeline

Parallelize

Bypass

Location

QoS Security Billing ACL IP QoS Billing

Location

Data-plane Control-plane

Location Location

IP QoS Security Billing ACL Security QoS Security Billing ACL IP QoS Security Billing ACL

IP Data

IP QoS Billing

Location

UL/DL data services available

Acceleration with State Replicas:

Pipeline the Data and Control

IP QoS Security Billing ACL

Location Location

In-band signaling installs states earlier on the data path

Pipeline

Parallelize

Bypass

IP QoS Security Billing ACL

Location Location

QoS Security Billing ACL IP QoS Billing Locatio n

Data-plane Control-plane

Location Location

IP QoS Security Billing ACL Security QoS Security Billing ACL IP QoS Security Billing ACL IP QoS Billing

Location

Concurrent operations via replicas

IP QoS Security Billing ACL

Location

Billing QoS IP QoS Security Billing ACL Billing Billing IP Security QoS QoS

Acceleration with State Replicas:

Parallelize the Control Procedures

Concurrent operations via replicas Mobility Controller has the final say

Location Location

Pipeline

Parallelize

Bypass

DPCM: a new perspective on data latency in LTE

We revisit control-plane functions in LTE from distributed state mgmt perspective Oh wait…

• How are transient failures handled?

✦ Replicas tolerates failures and prevents control-plane blocking

• What if concurrent state conflicts arise?

✦ Infrequent: 1.3% possible RW conflicts from operational LTE traces ✦ Fingerprint check + mobility controller final say -> no worse than LTE

25

Non-mandatory procedures block data Parallelizable procedures are forced to run sequentially Failures block the entire

• perations

Pipeline control procedures w/ data forwarding Parallelize multiple control procedures Bypass slow bottleneck control procedures

Implementation

A sohware extension of OpenAirInteiace (OAI) Incremental deployability and backward compatibility Network-side vi\ualization + device-side daemon

26

4G-LTE stack shim layer 4G-LTE stack shim layer

Extension daemon

4G-LTE stack

4G-LTE stack

4G-LTE stack

Legacy 4G device Supported device Data plane Control plane

shim layer

Base Station Mobility Controller Gateway

Evaluation

Testbed conggured to approximate real networks

• Network: 7 VMs running OAI+DPCM
• Device: OAI’s built-in emulator
• Parameters/latencies/failures from user study logs

Compare with 4G LTE in three dimensions

• Control-plane latency reduction
• Benefits for mobile applications
• System overhead

27

Control-plane latency reduction

28

Latency (ms)

250 500 750 1000 Static Mobility DPCM 4G LTE

2.1x 5.8x

4G LTE

Latency (ms)

250 500 750 1000 Static Mobility DPCM 5G (projection)

13.4x 88.9x

5G New Radio Projection (Assuming 1ms radio latency) More reductions in failure handling: up to 11s (11.5x in 4G, 317.8x in 5G) Greater latency reduction in mobility scenario Control plane latency dominates if radio latency greatly reduces

BeneUt on mobile apps

29

w/o failure w/ failure

Latency (s) 2.5 5 7.5 10

Web loading YouTube buffering

DPCM 4G LTE

Latency (s)

3.5 7 10.5 14

Web loading YouTube buffering

DPCM 4G LTE

1.7x 2.1x 2.1x 3.7x

Apps experience longer delay than data access latency due to imperfect adaptations

Acceptable system overhead

30

Overhead Network Node Mobile Device Signaling 56 bytes/msg 312 bytes/msg CPU 1.3% 0.1% Memory 53.9 MB 3.5 MB

Insights

Current LTE control plane operations are network-centric and sequential

• Sequentiality + network-centric is simply and natural engineering solution
• Correctness is probably a very important consideration

DPCM: a device-centric initiative benefits from distributed state management

• Break the traditional “smart-core, dumb-end” constraint in telecom
• Accelerate using parallelization, with good understanding of domain knowledge
• Corner cases are non-negligible! Need new/creative solution (Fast Paxos…) and engineering

tradeoff (how frequent?)

31

Towards next-gen low-latency mobile networks

Low-latency mobile networks are critical for future-proof designs (5G, self driving, …) Explore the paradigm shih to device-centric, veri3able mobile networking

• Device-centric: correctness guarantee, verifiable
• Verifiable: designer for networks, not plumber…

Example: multi-carrier access design in Google Project Fi

• Device-side solution providing better network access w/o infrastructure upgrade
• Apply new verifiable design to take full advantage of device-centric approach

32

Project Fi Fi SIM card User Space Hardware

T-Mobile Sprint U.S. Cellular Google Server (b) (a) (c) (d)

Conclusion

New source of latency for mobile apps — LTE control plane operations Our breakdown analysis fu\her deciphers the components and root causes We take a device-centric approach to speed up current LTE control-plane

• peration with new design [DPCM, MobiCom 2017]
• http://metro.cs.ucla.edu/dpcm.html
• Preliminary release on GitHub

Full possibilities for future low-latency network design!

33

Q & A