Post on 08-Jul-2018
QUALCOMM CONFIDENTIAL AND PROPRIETARY
Understanding Mobile Web Browser Performance
1
Rajiv VijayakumarQualcomm Incorporated
Topics
§Mobile browser architecture§Mobile network latencies§Android HTTP: Parallel downloads and caching§Power consumption
2
\3
SmartPhone Web Browsers
\4
What is a “WebKit” Browser?
Browser Architecture
5
Layout Engine
Browser Architecture
5
Browser Architecture
5
WebKit
Browser Architecture
5
WebKit JavaScriptEngine
Browser Architecture
5
WebKit JavaScriptCoreor V8
Browser Architecture
5
WebKit
Rendering Networking
JavaScriptCoreor V8
Browser Architecture
5
Browser UI
WebKit
Rendering Networking
JavaScriptCoreor V8
\6
Android
\7
Android Browser ≠ Chrome
<img src="http://www.qualcomm.com/imgA.jpg">
8
<img src="http://www.qualcomm.com/imgA.jpg">
8
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNS
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
ww
w.q
ualc
omm
.com
96.1
6.97
.9DNS
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNS TCP
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNSS
YN
SY
N/A
CK
TCP
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNS TCP HTTP
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNS TCP
GE
T /im
gA.jp
g
HTT
P/1
.1 2
00 O
K
HTTP
Time
<img src="http://www.qualcomm.com/imgA.jpg">
8
DNS TCP
GE
T /im
gA.jp
g
HTT
P/1
.1 2
00 O
K
HTTP
Time
Latencies in 3G/4G Networks
9
Latencies in 3G/4G Networks
§ Initial connection setup: > 1s (on older 3G)
9
Latencies in 3G/4G Networks
§ Initial connection setup: > 1s (on older 3G)§Round trip time (RTT): 80ms or higher (typically)§ 100 - 200ms common
9
RTTs at speed of light
§San Diego – New York : ~26ms
§San Diego – Johannesburg : ~110ms
10
Every round trip is at least 80-100ms
11
DNS TCP HTTP
Every round trip is at least 80-100ms
Including HTTP retrievals from a CDN
11
DNS TCP HTTP
DNS
0
50
100
150
Seconds0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Histogram of DNS Lookup Times
Source: Data collected over 30 sites (280 domains) on Qualcomm corporate network12
Every new lookup takes 1 RTT + Lookup Time
Domain Sharding
§Use of multiple domains (cdn1.domain.com, cdn2.domain.com)
§Requests may not happen in parallel...
13
Parallel Downloads
14
cdn1.domain.com/img1A.pngcdn1.domain.com/img1B.pngcdn1.domain.com/img1C.pngcdn1.domain.com/img1D.png
cdn2.domain.com/img2A.pngcdn2.domain.com/img2B.pngcdn2.domain.com/img2C.pngcdn2.domain.com/img2D.png
8 images across 2 domains
cdn1
img1A
cdn2
Browsers typically use 4-6 sockets per host
15
img1B
img1C
img1D
img2A
img2B
img2C
img2D
t=0
cdn1
img1A
cdn2
Android
16
img1B
img1C
img1D
img2A img2Bimg2C
img2D
cdn1
img1A
cdn2
Android
16
img1B
img1C
img1D
img2A img2Bimg2C
img2D
cdn1
img1A
cdn2
Android
17
img1B
img1C
img2A
cdn1
img1A
cdn2
Android
17
img1B
img1C
img2A
§ “Thread per socket” model
§4 HTTP threads
QUALCOMM CONFIDENTIAL AND PROPRIETARY18 Source: DDMS (from Android SDK) connected to Android 2.3.3 emulator
cdn1
img1A
cdn2
Android
19
img1B
img1C
img2A
§ “Thread per socket” model
§4 HTTP threads
cdn1
img1A
cdn2
Android
19
img1B
img1C
img2A
§ “Thread per socket” model
§4 HTTP threads
§Only 4 sockets can have outstanding HTTP requests (“active”)
Android
20
cdn2
img2A img2Bimg2C
img2D
cdn1
img1A
img1B
img1C
img1D
Android
20
cdn2
img2A img2Bimg2C
img2D
Android
20
cdn2
img2A img2Bimg2C
img2D
Pipelining
Android
20
cdn2
img2A img2Bimg2C
img2D
§Send multiple requests without waiting for each response§ Responses must arrive in order
Pipelining
Android
20
cdn2
img2A img2Bimg2C
img2D
§Send multiple requests without waiting for each response§ Responses must arrive in order
§Android pipelines up to 3 requests on each socket
Pipelining
§Not supported, or not enabled by default on most major browsers
§Historically caused problems (e.g. due to badly behaving proxies)
§Millions of Android devices actively using pipelining § Time to revisit... ?
Pipelining
21
HTTP Caching
22
HTTP Caching
§WebKit maintains a memory cache§ Default size = 8 MB
§Android has a persistent (file system) HTTP cache§ Default size = 6 MB
22
Cache Eviction
23
Cache Eviction
§What gets evicted to make room?
23
Cache Eviction
§What gets evicted to make room?
§Android policy: Prioritize base on expiration date§ Far future expiration dates have higher priority§ Evict lowest priority items first
23
Expires:Wed, 09 Jun 2021 21:12:36 GMT
Example of Cache ContentsVisited several sites, read blog posts and linked articles etc.
24
Example of Cache ContentsVisited several sites, read blog posts and linked articles etc.sqlite3 webviewCache.db 'select expires from cache order by expires desc;'
24
Example of Cache ContentsVisited several sites, read blog posts and linked articles etc.sqlite3 webviewCache.db 'select expires from cache order by expires desc;'
Expiration Year
2036
2035
...
2011
24
Example of Cache ContentsVisited several sites, read blog posts and linked articles etc.sqlite3 webviewCache.db 'select expires from cache order by expires desc;'
Expiration Year
2036
2035
...
2011
Cou
nt
0
50
100
150
200
Expiration Year2010 2020 2030 2040
24
§Cache eviction policy affects performance
§Consider real-world cache occupancy when measuring cached page load performance
25
Caveats
§Devices can (and do!) differ from stock Android
§Differences in:§ Number of HTTP threads§ Cache size and eviction algorithm§ etc.
26
PowerWhat is the power drawn from the battery while loading a web page?
27
28 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
Battery Current (mA)
QUALCOMM CONFIDENTIAL AND PROPRIETARY29 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
11s download (400mA)
10s idle (300mA)Baseline (200 mA)
Power§Any network traffic brings up the radio for several seconds§ Network dormancy timer: typically 10-15 seconds
§Use long periods for any periodic Ajax calls etc.
30 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
TCP FINs and Power
31 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
TCP FINs and Power§Android closes TCP sockets after ~6s of inactivity§ Sockets get closed at different times
31 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
TCP FINs and Power§Android closes TCP sockets after ~6s of inactivity§ Sockets get closed at different times
§Every socket close generates a TCP FIN§ Extends the dormancy timer§ Keeps the modem on for a 1-bit communication
31 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
TCP FINs and Power§Android closes TCP sockets after ~6s of inactivity§ Sockets get closed at different times
§Every socket close generates a TCP FIN§ Extends the dormancy timer§ Keeps the modem on for a 1-bit communication
§Optimization: close all sockets after page load completes
31 Source: Power measurements made on MDP MSM8655™ running Android 2.3 over commercial 3G network
JavaScript Performance
§SunSpider benchmark numbers have improved dramatically in recent years§ Android (Jun 2011) < 2.5s§ Desktop Chrome (Sep 2008) ~2.8 s
32 Source: http://ejohn.org/blog/javascript-performance-rundown/
developer.qualcomm.com/webtech
33
34
Q&A
Nothing in these materials is an offer to sell any of the components or devices referenced herein. Certain components for use in the U.S. are available only through licensed suppliers. Some components are not available for use in the U.S.
Disclaimer
35
QUALCOMM Incorporated5775 Morehouse Drive
San Diego, CA. 92121-1714U.S.A.
Copyright © 2011 Qualcomm Incorporated. All rights reserved.
Not to be used, copied, reproduced in whole or in part, nor its contents revealed in any manner to others without the express written permission of Qualcomm.
Qualcomm and MSM8655 are trademarks of Qualcomm Incorporated, registered in the United States.
Other product and brand names may be trademarks or registered trademarks of their respective owners.