1 PREFETCHING INLINES TO IMPROVE WEB SERVER LATENCY Ronald Dodge US Army [email protected] Daniel Menascé, Ph. D. George Mason University [email protected]

1

PREFETCHING INLINES TO IMPROVE WEB SERVER

LATENCY

Ronald DodgeUS Army

[email protected]

Daniel Menascé, Ph. D.George Mason [email protected]

Dodge & Menascé. All Rights Reserved.

2

Electronic Commerce: online sales are soaring

“… IT and electronic commerce can be expected to drive economic growth for many years to come.”

The Emerging Digital Economy,

US Dept. of Commerce, 1998.


3

Business in the Internet Age (Business Week, June 22, 1998)

Type of Business 1997 2001 (forecast)Business to Business 8.000 183.000Travel 0.654 7.400Financial Services 1.200 5.000PC Hardware & Software 0.863 3.800Entertainment 0.298 2.700Ticket Event Sales 0.079 2.000Books & Music 0.156 1.100Apparel & Footware 0.092 0.514Total 11.342 205.514

4

Caution Signs Along the Road

There will be jolts and delays along the way for electronic commerce: congestion is the most obvious challenge.

(Gross & Sager, Business Week, June 22, 1998, p. 166.)

5

Outline

• Response Time Reduction Techniques

• HTTP Transactions without Prefetching

• HTTP Transactions with Prefetching

• Model– Workload characterization– Simulation Model

• Results

• Concluding Remarks


6

Outline





• Results



7

Response Time Reduction Techniques

• Caching– web browser – proxy server

• Prefetching– predict what documents will be requested

and bring them into the cache ahead of time.


8

Browser Caching

Clients

ProxyServer External Web

Servers

router(50sec/packet)

Internet

LAN (10 Mbps Ethernet)

.

.

.


9

Browser Caching

Clients


Servers


Internet


.

.

.

1

2

3

4Dodge & Menascé. All Rights Reserved.

10

Browser Caching

Clients


Servers


Internet


.

.

.3

4

1

2


11

Browser Caching

Clients


Servers


Internet


.

.

.3

4

1

2


12

Caching Proxy Server

Clients


Servers


Internet


.

.

.

1

2

3


13


Clients


Servers


Internet


.

.

.

1

2

3


14


Clients


Servers


Internet


.

.

.

1

2

3


15


Clients


Servers


Internet


.

.

.

1

2

3


16


Clients


Servers


Internet


.

.

.

1

2

3


17


Clients


Servers


Internet


.

.

.

1

2

3


18

Goal

Improve user perceived latency through caching and

prefetching of inlines at the server side.


19

Related Work on Prefetching

• Padmanabhan, 1995– effects of prefetching on latency and network load.

• Chinen and Yamaguchi, 1996– prefetching to the proxy server the first N links of any

Web page.

• Crovella and Badford, 1997– network load can be minimized by adjusting prefetch

rate

• Foxwell and Menasce, 1998– prefetching results of search engine queries.


20

Outline





• Results



21

Browser ServerHTTP request

inline 1 request

inline 2 request

HTTP document

inline 1 file

inline 2 file

HTML documentparsed bythe browser

server disk

No Caching/Prefetching of Inlines


www.cs.gmu.edu

22

<H1><CENTER>Capacity Planning for Web Performance</H1><hr></CENTER><hr>

The introduction:<p><ul><dt><img src="http://www.cs.gmu.edu/dcc/icons/ball-blue.gif">[<a href="#intro">Introduction</a>]

<dt><img src="http://www.cs.gmu.edu/dcc/icons/ball-green.gif">[<a href="#goals">Capacity Planning</a>]

<dt><img src="http://www.cs.gmu.edu/dcc/icons/ball-red.gif">[<a href="#why economics">Web Performance Problems</a>]

<dt><img src="http://www.cs.gmu.edu/dcc/icons/ball-yellow.gif">[<a href="#questions">Web Performance Modeling</a>]

<dt><img src="http://www.cs.gmu.edu/dcc/icons/ball-orange.gif">[<a href="#papers">Publications</a>]</ul>


inline

23

Outline





• Results



24

Browser ServerHTTP request

inline 1 request

inline 2 request

HTTP document

inline 1 file

inline 2 file

HTML documentparsed bythe browser

server disk

HTML documentparsed byserver

cache

Caching/Prefetching of Inlines


www.cs.gmu.edu

25

Outline





• Results



26

Workload Characterization

Analysis of the HTTP log of a Web server at GMU:

- requester’s address- request arrival time- type of request- file requested- size of file requested


27

Previous Workload Characterization Work

Arlitt and Williamson (1996), Crovella and Bestravos (1996)

� The average size of a transferred document does not exceed 21KB

� Less than 3% of the requests are for distinct files.� The file size distribution is heavy-tailed. � File inter-reference times are exponentially

distributed and independent.


28

Interrarival time distribution

0.0

0.2

0.4

0.6

0.8

1.0

0 4 8 12 16 20 24 28 32 36

time (sec)

PDF cdf

F x eT

x~

.( ) 1 0 4493


average = 2 sec.

29

HTML file size distribution

0.0

0.2

0.4

0.6

0.8

1.0

0 5 10 15 20 25 30 35 40Size (Kbytes)

PDF cdf

724.1753.1)(~ xxF

H


average = 3,552 bytes.

30

Distribution of number of inlines per HTML document

0.0

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100

Number of files

PDF cdf

163.2748.1)(~ xxF

NDodge & Menascé. All Rights Reserved.

average = 2.92 inlines.

31

Inline file size distribution

0.0

0.2

0.4

0.6

0.8

1.0

0 50 100 150

size (Kbytes)PDF cdf

473.1754.1)(~ xxF

IDodge & Menascé. All Rights Reserved.

average = 10,392 bytes.

32

File InterreferenceA.html B.html C.html D.html A.html

interreference distance = 3

Repeatfile?

Use interreference distributionto decide how many requests to look

back.

YES

Pick a new fileNO


33

File interreference distribution

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

0 20 40 60 80 100 120

PDF cdf

6204.005.00002.)( 2~ xxxFF


34

Outline





• Results



35

Network

CPU

Web browsers

WEB Server Cache

Disk

h

1 - h

Simulation Environment:


36

Simulation Experiment

• Simulation Tool: CSim

• Batch means analysis: accuracy 10% and confidence level of 95%.

• Cases studied:

Server Cache Size (Kbytes):0 16 32 64 128 256 512 1,024Number of clients:5 10 15 20 25


37

Outline





• Results



38

Response time (in sec) of HTML document requests vs. cache size (in Kbytes)

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0 128 256 384 512 640 768 896 1024

5 Clients 10 Clients 15 Clients 20 Clients 25 Clients


39

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0 128 256 384 512 640 768 896 1024


Response Time of Inline Files (in sec) vs. Cache Size (KB)


40

Total Response time (in sec) vs. cache size (in Kbytes)

0.00.10.20.30.40.50.60.70.80.9

0 128 256 384 512 640 768 896 1024



41

File Hit Ratio vs. cache size (in Kbytes)

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 128 256 384 512 640 768 896 10245 Clients 10 Clients 15 Clients 20 Clients 25 Clients


42

Byte Hit Ratio vs. cache size (in Kbytes).

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 128 256 384 512 640 768 896 1024



43

Server disk utilization vs. cache size (in Kbytes).

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0 128 256 384 512 640 768 896 1024



44

Total Response Time (sec) vs Cache Size (KB) for Prefetching and non-Prefetching Cases

0

2

4

6

8

10

12

14

0 128 256 384 512 640 768 896 1024

prefetching non-prefetching


caching used in both cases

45

Outline





• Results



46

Concluding Remarks

• Prefetching can be used in conjunction with caching to decrease user perceived latency.

• No modification to the HTTP protocol nor the browser software is needed.

• Results showed a 48% improvement in response time through prefetching of inlines.


47

Relevant Bibliography• Menascé, D. A. and V. A. F. Almeida, Capacity

Planning for Web Performance: metrics, models, and methods,” Prentice Hall, Upper Saddle River, NJ, 1998.

• N. Yeager and R. McGrath, Web Server Technology, Morgan Kaufmann, San Francisco. 1996.

• Foxwell, H. and D. A. Menascé, Prefetching Results of Web Searches, Proc. of the 1998 Computer Measurement Group Conference, Anaheim, CA, Dec. 6-11, 1998.


Documents

1 PREFETCHING INLINES TO IMPROVE WEB SERVER LATENCY Ronald Dodge US Army [email protected] Daniel Menascé, Ph. D. George Mason University [email protected]