转http://blog.csdn.net/louieuser/article/details/6152175
Android应用的性能如何测试?JAVA层面可以用TraceView,可是用NDK开发出来的是so,TraceView跟踪不了怎么办?问了Google大神,答案是OProfile!
Oprofile 是Linux系统下一个低开销的系统全局的性能监视工具,利用处理器上所包含的专用的性能监视硬件(若没有性能监视硬件则使用一个基于计时器的代用品)来收集与性能相关的数据样品。它获得关于内核以及系统上的可执行文件的信息,例如内存是何时被引用的;L2缓存请求的数量;收到的硬件中断数量等。
Oprofile的特点如下:
l 无需重新编译源代码,如果不进行源代码及分析,连调试信息(-g option to gcc)也不是必须的。
l 只在内核中插入一个模块。
l 可以分析运行于系统之上的所有代码(禁用中断的代码除外)
l 系统的额外开销小,Oprofile会增加1%-8%的系统开销(取决于采样频率)
l 兼容所有2.2,2.4,2.6内核,可以运行在SMP系统之上
l 支持主流CPU架构,包括X86、arm、AVR32、mips、powerpc等
Oprofile要想跑在Andorid上,要满足下面的条件:
1.内核要支持
2.要将Oprofile移植到Arm平台上
下面是移植的全过程:
一、Oprofile移植
用到的交叉编译工具如下:
arm-2010.09-50-arm-none-linux-gnueabi-i686-pc-linux-gnu.tar.bz2
用到的库如下:
popt-1.14.tar.gz
binutils-2.21.tar.gz
oprofile-0.9.6.tar.gz
$ tar xvfz arm-2010.09-50-arm-none-linux-gnueabi-i686-pc-linux-gnu.tar.bz2 -C ~/
修改~/.bashrc,添加
export PATH=${PATH}:/home/louieli/arm-2010.09/bin
$ tar zxvf popt-1.14.tar.gz
$ cd popt-1.14
$ ac_cv_va_copy=yes ./configure --with-kernel-support --host=arm-none-linux-gnueabi --prefix=/home/louieli/work/popt
$ make
$ make install
$ tar zxvf binutils-2.21.tar.gz
$ cd binutils-2.21/
$ ./configure --with-kernel-support --host=arm-none-linux-gnueabi --prefix=/home/louieli/work/binutils --enable-shared
$ make LDFLAGS="-all-static"
可能会出现 cc1: warnings being treated as errors,找到出错文件的Makefile文件,将-Werror去掉
$ make install
$ tar zxvf oprofile-0.9.6.tar.gz
$ cd oprofile-0.9.6/
$ ./configure --with-kernel-support --host=arm-none-linux-gnueabi --prefix=/home/louieli/work/oprofile/ --with-extra-libs=/home/louieli/work/popt/lib/ --with-extra-includes=/home/louieli/work/popt/include/ --with-binutils=/home/louieli/work/binutils
$ make LDFLAGS="-all-static -L/home/louieli/work/binutils/lib -Xlinker -R -Xlinker /home/louieli/work/binutils/lib -L/home/louieli/work/popt/lib/"
$ make install
用file 命令查看,我们需要的oprofile文件都已经变成可以在android上跑的静态链接文件了
install.sh: Bourne-Again shell script text executable
opannotate: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
oparchive: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
opcontrol: a /system/bin/sh script text executable
opgprof: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
ophelp: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
opimport: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
opjitconv: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
opreport: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
oprofiled: ELF 32-bit LSB executable, ARM, version 1 (SYSV), statically linked, for GNU/Linux 2.6.16, not stripped
二、编译linux内核映像
a)准备交叉编译工具链
android代码树中有一个prebuilt项目,包含了我们编译内核所需的交叉编译工具。
b)设定环境变量
$ emacs ~/.bashrc
增加如下两行:
export PATH=$PATH:~/android/prebuilt/linux-x86/toolchain/arm-eabi-4.4.0/bin
export ARCH=arm
保存后,同步变化:
$ source ~/.bashrc
c)获得合适的内核源代码
$ cd ~/android
获得内核源代码仓库
$ git clone git://android.git.kernel.org/kernel/common.git kernel
$ cd kernel
$ git branch
显示
* android-2.6.27
说明你现在在android-2.6.27这个分支上,也是kernel/common.git的默认主分支。
显示所有head分支:
$ git branch -a
显示
* android-2.6.27
remotes/origin/HEAD -> origin/android-2.6.27
remotes/origin/android-2.6.25
remotes/origin/android-2.6.27
remotes/origin/android-2.6.29
remotes/origin/android-goldfish-2.6.27
remotes/origin/android-goldfish-2.6.29
我们选取最新的android-goldfish-2.6.29,其中goldfish是android的模拟器模拟的CPU。
$ git checkout -b android-goldfish-2.6.29 origin/android-goldfish-2.6.29
$ git branch
显示
android-2.6.27
* android-goldfish-2.6.29
我们已经工作在android-goldfish-2.6.29分支上了。
d)设定交叉编译参数
打开kernel目录下的Makefile文件,把CROSS_COMPILE指向刚才下载的prebuilt中的arm-eabi编译器
CROSS_COMPILE ?= arm-eabi-
把
LDFLAGS_BUILD_ID = $(patsubst -Wl$(comma)%,%,/
$(call ld-option, -Wl$(comma)–build-id,))
这一行注释掉,并且添加一个空的LDFLAGS_BUILD_ID定义,如下:
LDFLAGS_BUILD_ID =
e)编译内核映像
$ cd ~/android/kernel
$ make goldfish_defconfig
$ make menuconfig
修改内核配置如下
General setup --->
[*] Profiling support (EXPERIMENTAL)
[ ] Activate markers
[*] OProfile system profiling (EXPERIMENTAL)
这是把OProfile直接编进内核,也可以选择[M] OProfile system profiling (EXPERIMENTAL)会在arch/arm/oprofile文件夹下生成oprofile.ko,oprofile.ko需要用insmod载入。
$make
f)测试生成的内核映像
$ emulator -avd myavd -kernel ~/android/kernel/arch/arm/boot/zImage
三、Oprofile在android模拟器中的使用
1.先看一下opcontrol的参数
# opcontrol
opcontrol: usage:
-l/--list-events list event types and unit masks
-?/--help this message
-v/--version show version
--init loads the oprofile module and oprofilefs
--setup give setup arguments (may be omitted)
--status show configuration
--start-daemon start daemon without starting profiling
-s/--start start data collection
-d/--dump flush the collected profiling data
-t/--stop stop data collection
-h/--shutdown stop data collection and kill daemon
-V/--verbose[=all,sfile,arcs,samples,module,misc,ext]
be verbose in the daemon log
--reset clears out data from current session
--save=name save data from current session to session_name
--deinit unload the oprofile module and oprofilefs
-e/--event=eventspec
Choose an event. May be specified multiple times. Of the form
"default" or "name:count:unitmask:kernel:user", where :
name: event name, e.g. CPU_CLK_UNHALTED or RTC_INTERRUPTS
count: reset counter value e.g. 100000
unitmask: hardware unit mask e.g. 0x0f
kernel: whether to profile kernel: 0 or 1
user: whether to profile userspace: 0 or 1
-p/--separate=type,[types]
Separate profiles as follows :
none: no profile separation
library: separate shared library profiles per-application
kernel: same as library, plus kernel profiles
thread: per-thread/process profiles
cpu: per CPU profiles
all: all of the above
-c/--callgraph=#depth enable callgraph sample collection with a maximum depth.
Use 0 to disable callgraph profiling.
--session-dir=dir place sample database in dir instead of
default location (/var/lib/oprofile)
-i/--image=name[,names] list of binaries to profile (default is "all")
--vmlinux=file vmlinux kernel image
--no-vmlinux no kernel image (vmlinux) available
--kernel-range=start,end kernel range vma address in hexadecimal
--buffer-size=num kernel buffer size in sample units
--buffer-watershed kernel buffer watershed in sample units (2.6 only=
--cpu-buffer-size=num per-cpu buffer size in units (2.6 only)
--note-table-size kernel notes buffer size in notes units (2.4 only)
--xen Xen image (for Xen only)
--active-domains=<list> List of domains in profiling session (for Xen only)
(list contains domain ids separated by commas)
2.使用方法
将我们之前编译好的oprofile和busybox装入模拟器
执行oprofile目录中的install.sh 将oprofile装入模拟器
adb push busybox /data/busybox
$adb shell //进入模拟器shell
#chmod 777 /data/busybox
# /data/busybox --install /data/busybox
#export PATH=/data/busybox:$PATH:/data/oprofile
# mount -o remount rw /
# mount -o rw,remount -t yaffs2 /dev/mtdblock3 /system
# touch /etc/mtab
# echo nodev /dev/oprofile oprofilefs rw 0 0>/etc/mtab
# mkdir /dev/oprofile
# mount -t oprofilefs nodev /dev/oprofile //这一句很重要,没有这一句会出现下面的错误
# opcontrol --init
cat: can't open '/dev/oprofile/cpu_type': No such file or directory
Unable to open cpu_type file for reading
Make sure you have done opcontrol --init
cpu_type 'unset' is not valid
you should upgrade oprofile or force the use of timer mode
# opcontrol --init //初始化,只需运行一次
# opcontrol --setup --callgraph=2 --session-dir=/data/first --no-vmlinux
Using 2.6+ OProfile kernel interface.
Using log file /data/first/samples/oprofiled.log
Daemon started.
Profiler running.
# opcontrol --status
Daemon running: pid 637
Separate options: none
vmlinux file: none
Image filter: none
Call-graph depth: 2
# opcontrol --start //启动profiler
Using 2.6+ OProfile kernel interface.
Using log file /var/lib/oprofile/samples/oprofiled.log
Daemon started.
Profiler running.
# /data/test/test //运行我们的程序 ( 我的测试程序通过这条指令编译arm-none-linux-gnueabi-gcc -g -o test test.c -static -fno-omit-frame-pointer)
in c
in a
in b
in a
in c
in b
in a
in a
# opcontrol --dump //收集采样数据
# opcontrol --stop //停止profiler
Stopping profiling.
#opreport --session-dir=/data/first -l /data/test/test //查看报告
CPU: CPU with timer interrupt, speed 0 MHz (estimated)
Profiling through timer interrupt
samples % symbol name
11291 79.9589 a
1129 7.9952 b
853 6.0406 main
848 6.0052 c
现在我们就可以根据oprofile的输出对我们的程序进行优化了。
如果有哪位同学也想试一把的话,一定要用linux。这种移植环境很重要,我之前就在测试机(win7+cygwin)上浪费了很多时间。这里有打包好的工具,大家可以下载。其中kernel-qemu就是我们之前编译好的内核,替换掉Android SDK中的kernel-qemu就行了。祝各位好运!