Mooreâ€™s Law and Network Optimization

1

Moore’s Law and Network Optimization

Constantine D. PolychronopoulosUniversity of Illinois at Urbana-Champaign

Onassis Foundation Science Lecture Series 2008 Computer ScienceITE - CreteJuly 2008

2

Moore’s Law: Guiding IC and processor evolution for 40+ years

Capacity Speed (latency)Logic 2x in 3 years 2x in 3 yearsDRAM 4x in 3 years 2x in 10 yearsDisk 4x in 3 years 2x in 10 years

• Popular version: –Double CPU performance every 18 months

3

Tremendous impact on society• Mapping of human genome • Weather prediction and climate change

monitoring• Speech synthesis, AI applications, etc• Other Grand Challenge/NAE applications…

• … the above notwithstanding, what’s the impact of Moore’s Law (other than forcing us to spend a lot of money) on your and my experience using computers?

4

Moore’s Law versus User Experience• What is user experience?

The response/latency perceived by a user on a given application:

- web page download- file transfer, VOD- compilation task- IM - electronic games- VoIP- streaming media - …

5

Moore’s Law Paradox

User experience remains constant or improves marginally according to a step function

• Why? – Software layering & interoperability– Component, modular design => layers of libraries– Increasing application dimensioning– New, higher quality content (media) and interfaces– Complex HW, more difficult to fully utilize– Inefficient ways of writing software!

6

The Duality of Moore’s Paradox in Networking

Wireless broadband speeds double every three years, but Internet download times remain constant or improve marginally…

Average web page download time (broadband – wireline):- Feb’06: 2.8 sec- Feb’08: 2.35 sec

• Why? – The answers in numbers…

7

Broadband Internet - Performance Trends • Web pages are composed of XHTML container

objects (CO) and external objects (EO): images, video, audio, external CSS and JavaScript files

• Average web page has grown by 22x in the last 10 years – quadrupled from 2003-2007– Average page size close to 400KB– Average number of objects to 55/page– Top 1000 web pages grew by 26% 12/06-12/07

8

Key Metrics - 2007 (Cont’.)• JavaScript: 89% of web pages used the script

element:– Avrg no. of external scripts: 7– Avrg size of external script: 8.9KB– Total avrg script size: 69KB

• Use of CSS: 83% used the link tag and 55% used the style tag. Avrg size of external style sheets was 6.6KB. Total avrg style size was 15.2KB

• Use of images: 92% of web pages use them

9

Typical image objects in Web pages

Image Encoding Frequency 2006 Frequency 2007GIF 77.9% 84.6%JPEG 55.8% 64.5%PNG 7.2% 32.2%BMP 0.8%

Average aggregate graphic view area in a web page: 221X221 pixels[Gonzalez-Canete, Casilari, & Trivino-Cabrera, 2007]

10

Average Web page size: tripled 2003-2007

11

Size and number of objects growth

12

And the “bad” news…• Use of streaming media on the Web grew by

more than 100% each year from 2003-2007• In the period 2000-05 the total volume of

streaming media files stored on the Web grew by more than 600%

• More than 87% of all streaming media is abandoned by users within 10 seconds of session initiation (wasting more than 20% of server bandwidth)

• YouTube: Only 3% of server responses are for video, but account for over 98.6% of the bytes transferred

13

More video traffic metrics• Over 40% of business users experienced quality

degradation with videos over 30secs (re-buffering, stream switching, video cancelation etc).

• As broadband penetration increases so does video size, bit rate and duration.

• Median bit rate of web videos grew from 200Kbps in 2005 to 328Kbps on YouTube in 2007.

– Median video file size at over 63MB in 2007– Average YouTube video size was 10MB in ’07 with

65K new videos added every day

14

Video content growth: Average video object

Duration ≠ Size

15

Aggregate data traffic growth in WWANs• Europe and North America networks experience a

15-20% growth month over month or more than 3x compounded annually!

– What is even more impressive is the slop of growth continues to grow…i.e., exponential growth

• Current response: Ad hoc - no good solution– Block any non-HTTP traffic– Throttle downloads and streaming traffic to a crawl

• Projection: Over 80% of all Internet traffic will be streaming media by 2015 … or sooner?

16

What does this all mean?• Latency due to object overhead now dominates

most web page delays• Narrowband users (ISDN) are experiencing a

slowdown• Broadband users have experienced a marginal

improvement with average download time decreasing from 2.8sec in 2/06 to 2.35sec in 2/08 (KB40 – Keynote Business 40 Internet Performance Index) [Berkowitz & Gonzalez 2008]

• As video applications become of age, bandwidth will become a scarce resource

17

An Imperative for Network Optimization• We have invested in code optimization and

compilers since the early days of computers• Computational and memory optimization has

been the fabric of algorithms and complexity theory

• But how about (Wireless) Network Optimization?– Limited success in transport protocols (TCP)– Niche use of WWANs (circuit switched voice in

early cellular networks)– The world is changing rapidly and high

bandwidth “pipes” can’t keep up

18

The cost factor: Network Provisioning• Cost of wireless/cellular networks measures in

the $$ billions (spectrum, infrastructure, operations)

– Average of $0.20 per MB transported

• Provisioning around peak usage patterns• Over-provisioned networks are becoming

congested with the introduction of wireless broadband

– Dropped calls & connections– Long latencies– Unnecessary limitations on usage

19

Network Optimization• Affects every aspect of the cellular network

architecture: – Radio spectrum opimization (OFDM)– Physical link layer – IP layer– Transport protocol layer – Application layer

• Standards specifications bodies– IEEE, ITU, IETF, 3GPP

20

Abstraction of Wireless Wide Area Networks

Optimization function

21

Protocol Stacks – that was yesterday: toward all IP

22

Transparency of Optimization is Important

ServerClient

TCP

socketSocket

Interface

IP

TCP

socketSocket

Interface

IP

TCP

Socket

MobileCore

NetworkClient Gateway

Internet

Classifier

Interface

Server

Optimization

23

How Optimization Works: Bearer Layer Agnostic

Network Operator’s IP Services Core

$500M – $10B Infrastructure

Internet

24

384 Kbps Unoptimized

UMTS


25

384 Kbps Unoptimized

UMTS


1Mbps Optimized

26

1.8 Mbps Unoptimized

HSDPA


27

1.8 Mbps Unoptimized

Unoptimized Optimized

HSDPA


3.6Mbps Optimized

28

Example: Unoptimized Web Page Download

JavaScript 2

HTMLDNS

Web

Pag

e O

bjec

t Ret

rieva

l

Unoptimized

JavaScript 1DNS

Image 8

Image 7

Image 6

Image 5

Image 4

Image 3

Image 2

Image 1

Image 10

Image 9

Time– Page Done

29

JavaScript 2

JavaScript 1DNS

Image 8

Image 7

Image 6

Image 5

Image 4

Image 3

Image 2

Image 1

Image 10

Image 9

HTML

With Universal Application Compression,

DNS

Web

Pag

e O

bjec

t Ret

rieva

l

Universal Application Compression makes objects smaller…

JS 2

HTML

I 8

I 7

I 6

I 5

I 4

I 3

I 2

I 1

I 10

I 9

JS 1DNS

TimeUnoptimized– Page Done

30

With Dynamic Optimization

HTMLDNS

Web

Pag

e O

bjec

t Ret

rieva

l

JS 1DNS

JS 2

I 8

I 7

I 6

I 5

I 4

I 3

I 2

I 1

I 10

I 9

Compression

Smaller objects download faster…

… average speed-up of 2x (1.5x to 4x)

TimeUnoptimized– Page Done

31

Performance Gain

HTMLDNS

Web

Pag

e O

bjec

t Ret

rieva

l

Compression

Objects are retrieved in parallel and don’t need to be in order…

JS 2

JS 1DNS

I 8

I 7

I 6

I 5

I 3

I 2

I 1

I 10

I 9

MDI-1

I 4

TimeUnoptimized

… average speed-up of 3x (2x to 7x)

– Page Done

32

With Object Prediction

HTMLDNS

Web

Pag

e O

bjec

t Ret

rieva

l

Compression

Client-Based Object Prediction: Objects are requested and retrieved earlier…

MDI-1

JS 1DNS

I 8

I 7

I 6

I 5

I 4

I 3

I 2

I 1

I 10

I 9

JS 2

MDI-2Time

Unoptimized

… average speed-up of 4x (2.5x to 10x)

– Page Done

33

High-Speed InternetWireless Network

LocalServer

CentralServer

AdServer

Step 1: Set up the connection

In practice, CNN has more than 75 objects on 9 servers.For simplicity, here we show only 12 objectson 3 servers.

Image

Image

Image

Image

Image

Image

Image

Image Image Image

Image

Image

Image

Standard HTTP in Wireless: First Access

34

High-Speed InternetWireless Network

Standard HTTP in Wireless: First Access

LocalServer

CentralServer

AdServer

Step 1: Set up the connection

In practice, CNN has more than 75 objects on 9 servers.For simplicity, here we show only 12 objectson 3 servers.

Image

Image

Image

Image

Image

Image

Image

Image Image Image

Image

Image

Image

Image

Image

Image

Image

Image

Image

Image

Image Image Image

Image

Image

Image

ImageImage

Image

Image

Image

Image

Image

ImageImage

Image Image Image

Image

35

High-speed InternetWireless Network

LocalServer

CentralServer

AdServer

OSNOSN

With Dynamic Interleaving, web sites download 3 to 10 times faster

Image

Image

Image

Image

Image

Image

Image

Image Image Image

Image

Image

Image

Image

Image

Image

Image

Image

Image

Image

Image Image Image

Image

Image

Image

ImageImage

Image

Image

Image

Image

Image

ImageImage

Image Image Image

Image

Dynamic Interleaving

36

Sources and Effect of Clientless Optimization

Local cacheJavaScript/CSS in-liningIncreased number ofsimultaneous HTTP requestsMulti-part optimization

End-to-End Latency Reduction Data Reduction

• Gzip compression of HTML• Lossy image reduction (and

other multimedia types)• Remove non-functional data

from HTML/Java-scripts/CSS• Aggressive caching in

browser (reduce latency too)

37

Performance Improvements

20-Second Bar

0

20

40

60

80

100

120

Banking Travel Shopping News Info Sports

Unoptimized Optimized

Seconds

Type of Web Site Downloaded

38

0

1

2

3

4

5

6

7

8

9

10


Speedup Factor


Performance Improvements

39

Increase in Network Efficiency

72%

74%

76%

78%

80%

82%

84%

86%

88%

90%

92%


Data Reduction


40

Reduction of 3G Network LatencySeconds Until Done


0

10

20

30

40

50

60

Gaming News Banking Reference Portal

Unoptimized

Optimized

41

QUANTITATIVE QUALITATIVE

Expanding network capacity

Improving user experience – enabling new applications

The many faces of Optimization

PerformanceEnhancements

LatencyReduction

BandwidthImprovement

TransportProtocol

Optimization

ApplicationProtocol

Optimization

ContentCaching

AdaptiveContent-awareCompression

IntelligentContent

Reduction

Accessibility

ContentTranscoding

ProtocolTranslation Usability

PolicyEnforcement QoS Content

AdaptationTraffic

Shaping

OPTIMIZATION

42

On-the-fly Multimedia ReductionSignificant data reductionLow delay/latencyTransparent

43

Dynamic Bandwidth Shaping

Media function constantly monitors the network connection with the client and shapes the multimedia stream to adapt to current network conditions.

Ensures uninterrupted streams at optimal rates.

44

Optimization as a catalyst for rich applications

Optimization provides perpetual value – more important in 4G: • No “fast” is “fast enough” – otherwise, no need for 3G, 4G,…• Optimization is as much about user experience as net utilization• In a congested network, user experience deteriorates exponentially This is

independent of 3G, 4G or beyond• 3G and 4G address bandwidth, but not latency• New services require ever increasing bandwidth and ever decreasing

latency (real-time applications: video conferencing, network gaming, etc)• Optimization addresses multiple dimensions of performance:

– Bandwidth and latency– Network congestion– Protocol and application inefficiencies– Device CPU and memory limitations– Content packaging– Access enablement– Radio Access Network agnostic optimization!

Mooreâ€™s Law and Network Optimization

Documents

Transcript of Mooreâ€™s Law and Network Optimization