Improve Android System Component Performance

Improve Android System
Component Performance

Jim Huang ( 黃敬群 ) <jserv@0xlab.org>
Developer & Co-Founder, 0xlab
http://0xlab.org/

Feb 14, 2012 / Android Builders Summit

Rights to copy
© Copyright 2012 0xlab
http://0xlab.org/
contact@0xlab.org
Attribution – ShareAlike 3.0
You are free Corrections, suggestions, and contributions are
welcome!
to copy, distribute, display, and perform the work
to make derivative works Latest update: Feb 14, 2012
to make commercial use of the work
Under the following conditions
Attribution. You must give the original author credit.
Share Alike. If you alter, transform, or build upon this work, you may distribute the
resulting work only under a license identical to this one.
For any reuse or distribution, you must make clear to others the license terms of this
work.
Any of these conditions can be waived if you get permission from the copyright holder.
Your fair use and other rights are in no way affected by the above.
License text: http://creativecommons.org/licenses/by-sa/3.0/legalcode

0x1ab = 162 + 16x10 + 11 = 427
(founded on April 27, 2009)

0xlab is another Hexspeak.
(pronounce: zero-aks-lab)

About Me (1) Come from Taiwan
(2) Contributor of Android
Open Source Project (AOSP)
(3) Developer, Linaro
(4) Focus: system performance
and virtualization at 0xlab

Mission of 0xlab development:
Improve UX in SoC

UX = User Experience
SoC = Integrated Computing Anywhere

Strategy and Policy

• open source efforts to improve AOSP
• We focus on small-but-important area of Android.
– toolchain, libc, dynamic linker, skia, GLES,
system libraries, HAL
• Develop system utilities for Android
– benchmark, black-box testing tool, validation
infrastructure
• Value-added features
– Faster boot/startup time, Bluetooth profile, visual
enhancements
• Submit and share changes to community
– AOSP, CyanogenMod, Android-x86
– Linaro

Working Model
Rowboat CyanogenMod Android-x86

Hidden Bugs in AOSP

• AOSP is dedicated to mobile
Lucky!
product devices shipped by OHA
members
We– Fixed hardware "bug" in Android accidently
encountered the and
specifications
– Not well verified for other
configurations
• Performance is important, but we
frequently hit the hidden bugs
when apply aggressive
optimizations.
– Quality is the first priority!

Quality in custom Android Distribution

• 0xlab delivers the advantages of open source
software and development.
– Quality relies on two factors: continuous
development + strong user feedback
• Several utilities are developed to ensure the
quality and released as open source software.
– 0xbench (Android benchmarking tool)
– ASTER (Android System Testing Environment and Runtime)
– LAVA (Linaro Automated Validation Architecture)
• In the meanwhile, performance is improved by
several patches against essential components.

Tip: Automate system before optimizing

LAVA: Automated Validation Infrastructure for Android

Android benchmark running on LAVA. Android support on LAVA
Automated Validation flow includes https://wiki.linaro.org/Platform/Validation/LAVA
from deploy, then reboot, testing,
benchmark running, and result submit. Android related commands in LAVA:
* deploy_linaro_android_image
* boot_linaro_android_image
* test_android_basic
* test_android_monkey
* test_android_0xbench
* submit_results_on_host

Check "LAVA Project Update"
by Paul Larson,
2012 Embedded Linux Conference

0xbench: comprehensive open source benchmark
suite for Android
• A set of system utilities for
Android to perform
comprehensive system
benchmarking
• Dalvik VM performance
• OpenGL|ES performance
• Android Graphics framework
performance
• I/O performance
• JavaScript engine performance
• Connectivity performance
• Micro-benchmark: stanard C library,
system call, latency, Java
invocation, ...

Project page: http://code.google.com/p/0xbench/

Collect and Analyze results on
server-side

Android Functional Testing

• stress test
– Utilizing 'monkey', which is part of framework
• Automated test
– Both blackbox-test and whitebox-test are required

Stress Test

• According to CDD (Compatibility Definition Document),
Device implementations MUST include the Monkey
framework, and make it available for applications to
use.
• monkey is a command that can directly talks to
Android framework and emulate random user input.
adb shell monkey p your.package.name v 500

• Decide the percentage of touch events, keybord
events, etc., then run automatically.

ASTER: Automated Test

• Blackbox-test vs. Whitebox-test
• An easy to use automated testing tool with IDE
– Built upon MoneyRunner
• Batch execution of visual test scripts
• Multiple chains of recall commands
• Designed for non-programmer or Q&A engineers
• Use OpenCV to recognize icons or UI hints

Project page: http://code.google.com/p/aster/

Improve Android System Component Performance

It is time to improve the
performance of Android system
components

No Silver Bullet
to Improve the whole

Possibly Premature optimizations in Android

• “Premature optimization is the root of all evil”
– Donald Knuth
• bionic libc
– glibc incompatibility, No SysV IPC, partial Pthread,
incomplete prelink
– inactive/incorrect kernel header inclusion
– May not re-use existing system utilities
• Assumed UI behavior
– Input event dispatching and handler
– Strict / non-maintainable state machine (policy)
– Depending on a certain set of peripherals
• Unclear HAL design and interface
– Wifi, Bluetooth, GPS, ...

Think Difficult

• To make performance improvement visible
– Modifications from Application level, Android
framework, system libraries, and kernel
• Slowdown in newer Android version
– Example: Graphics in Eclair (2.0/2.1) is much
slower than 1.5 or 1.6
• To optimize or not to optimize, that is the question.
– Merge Local optimizations != Optimized
globally
– Many Android applications don't take various
devices into consideration. Thus, performance
issues occur all the way.

Which parts will be Improved?

• 2D/3D Graphics
• Android Runtime
• Boot time

Three frequently mentioned items in Android engineering are selected
as the entry points: 2D/3D graphics, runtime, and boot time.

Functional View (1.5)
Applications
Applications
Gallery Phone Web Browser Google Maps ・・・・・

Android Framework
Android Framework
Activity Window Content Notification
View System
Manager Manager Manager Manager

Package Telephony Resource Location

System Library
System Library Android Runtime
Android Runtime
SurfaceFlinger
SurfaceFlinger OpenCORE
OpenCORE SQLite Class Library

Dalvik Virtual Machine
OpenGL|ES
OpenGL|ES AudioFlinger
AudioFlinger WebKit

SGL
SGL OpenSSL bionic libc
bionic libc Freetype

Linux Kernel
Linux Kernel

Functional View (2.3)
Applications
Applications
Gallery Phone Web Browser Google Maps ・・・・・

Android Framework
Android Framework
Activity Window Content Notification
View System

Package Telephony Resource Location
Manager Manager Manager Manager RenderScript

OpenGL|ES 2.x accelerated. System Library
System Library Android Runtime
Android Runtime
Drop 2D accel
SurfaceFlinger
SurfaceFlinger StageFright SQLite Class Library
V8 bridge Dalvik Virtual Machine
GLES 2.0 OpenGL|ES
OpenGL|ES AudioFlinger
AudioFlinger WebKit
SMP improvements
Skia
Skia OpenSSL bionic libc
bionic libc Freetype
JIT compiler
Skia supports GPU backend SMP fixes
In Android 3.x Linux Kernel
Linux Kernel

 Properties
Android SurfaceFlinger
 Can combine 2D/3D surfaces and surfaces from multiple applications
 Surfaces passed as buffers via Binder IPC calls
 Can use OpenGL ES and 2D hardware accelerator for its compositions
 Double-buffering using page-flip

from EGL to SurfaceFlinger

hgl = hardware
hgl = hardware agl = android software
agl = android software
OpenGL|ES
OpenGL|ES OpenGL|ES renderer
OpenGL|ES renderer

Android Graphics without OpenGL|ES
Hardware
Android Framework (Java)

EventHub libandroid_runtime
Surfaceflinger Copybit
(service) (HW accelerated)
Renamed to libgui
Renamed to libgui
in Android 4.0
in Android 4.0
libagl is an optimized GLES 1.x
libagl is an optimized GLES 1.x
libui
libGLES Impl. Android 4.0 comes with libAgl2,
Impl. Android 4.0 comes with libAgl2,
(libagl) which provides software GL ES 2.0
which provides software GL ES 2.0
Implementation using Pixelflinger2
Implementation using Pixelflinger2

libpixelflinger
libpixelflinger is software renderer
libpixelflinger is software renderer
Android 4.0 comes with a new implementation,
Android 4.0 comes with a new implementation,
When GLES doesn't work,
When GLES doesn't work, PixelFlinger2, which Is based on LLVM and
PixelFlinger2, which Is based on LLVM and
software is used Mesa (glsl2-llvm): external/mesa3d
Mesa (glsl2-llvm): external/mesa3d
software is used

2D Accelerator for Android Graphics
• libcopybit provides hareware bitblit operations which
includes moving, scaling, rotation, mirroring, and more
effects, like blending, dithering, bluring, etc.
• Removed since Android 2.3
– But adding it back might improve UX in large screen.
• Android has two copybit interfaces:
– Blit: moving / blending
– Stretch: scaling besides moving
• libcopybit is called by libagl which can do swapBuffers to
do the framebuffer page flipping that can also be
accelerated by libcopybit.

Copybit could improve the performance of page flipping
Copybit could improve the performance of page flipping

Copybit operations
Copybit: 2D blitter
Copybit: 2D blitter

Optimizing Graphics without 3D/HW

• Implement copybit HAL carefully
– Minimize clip region
– Eliminate data copy
• Check ioctl for page flipping in framebuffer driver
– Efficiency and consistency
• Without 3D/HW, Android Graphics is CPU bound
– Reduce the amount of surfaces to manipulate
– Optimizing skia (2D vector library) is important
– Optimize color space conversion
– Optimize blitter and primitive operations like
matrix using ARM VFP and NEON

Apply extra performance tweaks against optimized build
2D on Nexus S
(NEON)

2D Improvement (1)
external/skia/
ccommit ae265ac7f132f5d475040edf134e312b3987eade
Add NEON optimized blitter: RGB565 to ABGR8888 without filter
and blending

commit 4b9b68bb9b8f82d6f70d98449851bc4bb19958bd
optimize blend32_16_row and unroll SkRGB16_Blitter::blitRect

Reference benchmark using 0xbench 2D on Nexus S (1 GHz)
[before]
Draw Rect: 28.52 fps

[after]
Draw Rect: 37.89 fps

This presentation takes the contributions in CyanogenMod
as example including SHA-1 hash

2D Improvement (2)
external/skia/
commit cb837750a37d59c979768320a7cf5ced96c7231c
Add NEON optimized SkARGB32_Black_Blitter::blitMask

Reference benchmark results on Nexus S (ARM Cortex-A8; 1 GHz) using
skia_bench: (time in ms, smaller is better)
[before]
running bench [640 480] text_48_linear_pos
8888: cmsecs = 88.18
565: cmsecs = 61.51
running bench [640 480] text_48_linear
8888: cmsecs = 85.85
565: cmsecs = 60.18

[after]
running bench [640 480] text_48_linear_pos
8888: cmsecs = 38.52
565: cmsecs = 59.11
running bench [640 480] text_48_linear
8888: cmsecs = 36.24
565: cmsecs = 57.37
•

Benchmark: 3D (arm11-custom; no GPU)

This explains that we have several system tools and development flow
to help customers/community to verify the performance and improve.

Optimizing Graphics with 3D/HW

• The significant changes happen in applications and
Android (Java) framework usage
http://developer.android.com/guide/practices/design/performance.html

• Implement libgralloc carefully
– Minimize the overhead of graphics memory
allocator: the kernel helper
– Example: UMP (Unified Memory Provider) in ARM
Mali GPU
• Track the transactions inside SurfaceFlinger
– Eliminate the invalid layer operations
– Corresponding modifications in upper framework
• Still, page flipping benefits from libcopybit
– but it has smaller difference with 3D/HW

Arithmetic on Nexus S Tune Dalvik VM performance (armv7)

Arithmetic Improvements
• Floating-point performance depends on Dalvik VM.
• Internally, Dalvik VM has huge amount of byte-swapped access,
which can be improved by ARMv6's REV and REV16 instructions.
bionic/
commit 02bee5724266c447fc4699c00e70d2cd0c19f6e1
Use ARMv6 instruction for handling byte order

ARMv6 ISA has several instructions to handle data in different
byte order.

libcore/
commit 7d5299b162863ea898dd863004afe79f7a93fbce
Optimize byte-swapped accesses.

Brings the performance of byte-swapped accesses way down from about
3x to less than 2x worst-case (char/short) and 20% best-case
(long/double). The main active ingredients are switching to a
single-pass swapped-copy (rather than copy in one pass, swap
in a second pass), and ensuring we use ARM's REV and REV16
instructions.

bionic libc

• Android C/C++ library
• 0xlab/Linaro Optimizations (merged in AOSP)
– Memory operations: Use ARMv6 unaligned access to
optimize usual cases
• Useful to TCP/IP (big-endian ↔ little endian)
– Various ARM optimized functions
• memcpy, strcmp, strcpy, memset, memcpy, strlen
• sha1
• code size reduction: useful for recovery image

Prelinking in GNU world
(Quote from Embedded Linux optimizations – Size, RAM, speed,
power, cost by Michael Opdenacker
Thomas Petazzoni, Free Electrons)
• prelink
http://people.redhat.com/jakub/prelink/
• prelink modifies executables and shared libraries to
simplify the dynamic linker relocation work.
• This can greatly reduce startup time for big applications
(50% less for KDE!). This also saves memory consumed
by relocations.
• Can be used to reduce the startup time of a Linux system.
• Just needs to be run again when libraries or executables
are updated.
Details on http://elinux.org/Pre_Linking

Dynamic Linker Optimization:
Why and How?
• The major reason to optimize dynamic linker is to
speed up application startup time.
• Approaches:
● Implement GNU style hash support for bionic

linker
● Prelinker improvements: incremental global

prelinking
– reduce the number of ELF symbol lookup
aggressively
• Changed parts
– apriori, soslim, linker, elfcopy, elfutils

(normalized) Dynamic Link time

1
0.9
0.8
0.7
0.6
lp
0.5
gp
0.4 re.gp
0.3 re.pe.gp
0.2 re.pe.pgp.gp
re.pe.pgp.gp.are
0.1
0
ation

ess
ver

rild

lld
on

ggerd

r
ore

nage

insta
-daem
iaser

proc

keyst
anim

debu

cema
app_
med

dbus
boot

servi

(normalized) Symbol Lookup number

1

0.8

0.6
elf.lp
elf.gp
0.4 elf.re.gp
elf.re.pe.gp
0.2 elf.re.pe.pgp.gp
elf.re.pe.pgp.gp.are

0
on
ess

ggerd

rild

lld
ation

ore
ver

r
nage

insta
-daem
iaser

proc

keyst
anim

-0.2
debu

cema
app_
med

dbus
boot

servi

• DT_GNU_HASH: visible dynamic linking improvement =
Better hash function (few collisions)
+ Drop unnecessary entry from hash
+ Bloom filter

void foo (){
libc.so printf(“fooooo”);
printf bar();
}

libfoo.so
foo
bar

libfoo.so
DT_GNU_HASH DT_HASH
foo foo
bar bar
printf

Symbols lookup# fail# gnu hash filtered by bloom
in ELF
gnu.gp 3758 23702 19950 23310 18234 (78%)

gnu.gp.re 3758 20544 16792 19604 14752 (75%)

gnu.lp 61750 460996 399252 450074 345032 (76%)

gnu.lp.re 61750 481626 419882 448492 342378 (76%)

H = {x, y, z} = hash functions
Hash function may collision
→ Bloom filter may got false positives
Bit array

NOTE: Android 4.0 removes the support of prelinker,
but gnu style hash is still useful.

Case Study: WebKit in Android

event
WebCore Refresh the surface
(expose event)

Android.webkit.WebViewCore
android.webkit.WebView
Skia bridge
Skia bridge ...

WebKit
JNI
JNI
v8
v8
skia

Surface

How to Measure On Android/ARM?
• for Native libraries →
• Use 'perf' built without libperl, libpython
• oprofiled and opcontrol are there, CPU data is
missing
• Binaries for ARM need frame pointers to have
backtraces
• Java part is the performance hell always.
• traceview is a great tool for Java performance
analysis.
• JVMTI / JDWP (Java Debug Wire Protocol, normally
spoken between a VM and a debugger)

# Overhead Command Shared Object Symbol
# ........ ............... ..................... ......
#
89.23% system_server 2b0c6c [.] 0x000000002b0c6c
1.26% MLVdo_thread [kernel_helper] [k] 0x0000000017aa90
1.05% d.process.acore libskia.so [.] S32A_Opaque_BlitRow32_arm
0.83% d.process.acore libcutils.so [.] android_memset32
0.63% system_server libc.so [.] memcpy
0.63% d.process.acore libc.so [.] memset

system_server is the process name of Android
Framework runtime. It occupies most of CPU
resources, but it is hard to figure out details
only by native tools like perf.

We can always optimize known performance hotspot
routines such as S32A_Opaque_BlitRow32_arm but
should be measured in advance.

Approaches to Optimize WebKit

• Cherry-pick upstream enhancements
– Example: ARM NEON optimized renderer and blur
effects
• Track JNI bridge in WebKit – Avoid memory leaks
• Use hardware accelerated backing store for certain
UI actions such as scrolling
– Check Qualcomm's QAEP
• Image caching in both skia and webkit
• Since skia supports GL backend, webkit can utilize
the accelerated paths
– That's what Android 4.0 emphasize on.

Case Study: Profiling JNI

• Aprof : an Android profiler (by 0xlab, android-
platform@ mailing-list)
– a profiling tool for Android native code; aprof is not
only another gprof implement on Android but also
support for profiling shared
• The capability of aprof is similar to what gprof does, it
provides call graph and time sampling profiling, but
it's incompatible with gprof since the gprof can not
profile shared library.
– Limited by its representation and the fact of bionic
libc incompatibility with GNU world.
• Integrated with Android activity life-cycle

Aprof

% cumulative self self total
time time time calls ms/call ms/call name
99.52 2170 2140 2178309 0 0 fib
0.00 2170 0 1 0 217 main
0.48 0 30 0 0 0 <libc.so>

Android.mk
LOCAL_ENABLE_APROF := true

Android Boot Time Optimizations

Reducing Boot-Time is Art

• You have to take every piece of boot flow into
consideration.
• Linux Kernel itself usually contributes less time than
userspace.

Bootchart of Android 4.0 on Nexus S
We will focus on reducing “cold” boot time,
from power on to the execution of the system application.

Write Tiny Boot loader to Speed up
Qi U-Boot + XLoader
Boot-oader
Size ~30K ~270K+20K
Qi Boot-loader Time to Kernel <1s > 5s

 Only one stage boot-loader Usage Product Engineering

Code Simple Complicated
 Small footprint ~30 KB
 Currently support
− Freescale iMX31 Romcode
ROM Romcode
ROM
− Samsung S3C24xx
− Beagleboard
XLoader Qi
 KISS concept
− Boot device and load kernel
− 3 second reduction! U-boot
Uboot Linux

Linux
TI OMAP3

Optimized ARM Hibernation
• Based on existing technologies and little
modifications to userspace are required
– TuxOnIce
• Release clean-pages before suspend
• Swap out dirty-pages before save image
• Image size reduced leads to faster resume time.

Demo video: http://www.youtube.com/watch?v=pvcQiiikJDU
Beagleboard-xM (OMAP3)
Full source tree: http://gitorious.org/0xlab-kernel

Further Boot Time Optimizations
• Save the heap image (like core dump) of Zygote
after preloading classes
• Modify Dalvik to make hibernation image after
system init and before Launcher startup
• Parallize Android init
• Cache & share JIT'ed code fragment

Reference: File-Based Sharing For Dynamically Compiled Code
On Dalvik Virtual Machine, National Chiao Tung University in Taiwan

Improper Ethernet
bring-up blocking

Initial bootchart analysis:
Initial bootchart analysis:
(1) It takes 27s from HW reset to Android Launcher screen.
(1) It takes 27s from HW reset to Android Launcher screen.
(2) There is an improper Ethernet bring-up blocking for 2s.
(2) There is an improper Ethernet bring-up blocking for 2s.
(3) CPU usage looks busy.
(3) CPU usage looks busy.

Reduced from 27s to 22s
Android Launcher appears
earlier then previous scenario.

Remove “preloaded-classes" to
Remove “preloaded-classes" to
eliminate the time cost from Zygote
eliminate the time cost from Zygote
Remove unnecessary dependency
Remove unnecessary dependency Risk: potentially slower Android
Risk: potentially slower Android
to active services concurrently
to active services concurrently activity launch time
activity launch time

Reduce boot time without Hibernation

• Zygote (init2) takes a long time to initialize Dalvik
VM and Android framework, which are usually of
the same context in virtual memory view
• If we can capture the state of a running process in
Linux and save it to a file. This file can then be used
to resume the process later on, either after a reboot
or even on another machine.
http://cryopid.berlios.de/
https://ftg.lbl.gov/projects/CheckpointRestart/
http://dmtcp.sourceforge.net/
• Only not zygote can benefit from from process
freezing technique, but also system robustness
might be improved.

Conclusion

• Optimizing Android requires the collaboration from
community – verification, utilities, and upsteam
• UX is not as simple as its length.
– Always Do measurement before taking actions
– Hacking around the software stack
• Automated testing + continuous integration is
really important.

http://0xlab.org

Improve Android System Component Performance

More Related Content

What's hot (20)

Similar to Improve Android System Component Performance (20)

More from National Cheng Kung University (20)

Recently uploaded (20)

Improve Android System Component Performance