monkey 压力测试android
Monkey的特征
测试的对象仅为应用程序包,有一定的局限性。
Monkey测试使用的事件流数据流是随机的,不能进行自定义。
可对MonkeyTest的对象,事件数量,类型,频率等进行设置。
Monkey的基本用法
基本语法如下:
$ adb shell monkey [options]
如果不指定options,Monkey将以无反馈模式启动,并把事件任意发送到安装在目标环境中的全部包。下面是一个更为典型的命令行示例,它启动指定的应用程序,并向其发送500个伪随机事件:
$ adb shell monkey -p your.package.name -v 500
分析日志
通过Android trace文件分析死锁ANR实例过程
system/build.prop 日志文件主要记录手机系统信息,如版本,型号,品牌
adb logcat 导出日志文件
monkey.ini 配置文件
cmd=adb shell monkey -p com.dgm.user --throttle 500 --ignore-timeouts --ignore-crashes --monitor-native-crashes -v -v package_name=com.dgm.user logdir=d:\android remote_path=d:\android_server phone_msg_log=d:\android_temp\phone.txt sum = 100 - activity = com.dgm.user.SplashActivity exceptions=['NullPointer','IllegalState','IllegalArgument','ArrayIndexOutOfBounds','RuntimeException','SecurityException']
throttle 每次事件等待500毫秒
sum 定义随机事件数
exceptions 异常定义,用于后面扩展
结果生成为可视化图片 使用的是matplotlib
Paste_Image.png
当然可以看日志文件
代码分析
获得cpu-men
# -*- coding: utf-8 -*-import subprocess pkg_name = "com.dgm.user"cpu = [] men = []def top_cpu(pkg_name): cmd = "adb shell dumpsys cpuinfo | grep " + pkg_name temp = [] # cmd = "adb shell top -n %s -s cpu | grep %s$" %(str(times), pkg_name) top_info = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE).stdout.readlines() for info in top_info: temp.append(info.split()[2].decode()) # bytes转换为string # print("cpu占用:%s" %cpu) for i in temp: if i != "0%": cpu.append(i.split("%")[0]) return cpudef get_men(pkg_name): cmd = "adb shell dumpsys meminfo %s" %(pkg_name) print(cmd) temp = [] m = [] men_s = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE).stdout.readlines() for info in men_s: temp.append(info.split()) # print("内存占用:%s" %men[19][1].decode()+"K") m.append(temp) for t in m: men.append(t[19][1].decode()) return men
入口代码
import monkeyConfigfrom adb_common import AndroidDebugBridge as aiimport matplotlibBase as mtimport MenCpu as mimport datetime as dt CPU = [[],[]] # time,使用情况MEN = [[],[]] #当前时间,和内存使用情况# 得到手机信息def getPhoneMsg(cmd_log): l_list = [] f = open(cmd_log, "r") lines = f.readlines() for line in lines: line = line.split('=') #Android 系统,如anroid 4.0 if (line[0] == 'ro.build.version.release'): l_list.append(line[1]) #手机名字 if (line[0]=='ro.product.model'): l_list.append(line[1]) #手机品牌 if (line[0]=='ro.product.brand'): l_list.append(line[1]) f.close() return l_list#开始脚本测试def start_monkey(cmd, logdir, now1, logcatname): print(cmd) os.popen(cmd) # os.kill() #print"使用Logcat导出日志" cmd2 = "adb logcat -d >%s" % logcatname os.popen(cmd2) #print"导出traces文件" tracesname = logdir + "\\" + now1 + r"traces.log" cmd3 = "adb shell cat /data/anr/traces.txt>%s" % tracesname os.popen(cmd3)if __name__ == '__main__': ini_file = 'monkey.ini' if os.path.isfile(ini_file): if ai().attached_devices(): mc = monkeyConfig.baseReadnin(ini_file) ai().open_app(mc.get_package_name(), mc.get_activity()) os.system('adb shell cat /system/build.prop >'+mc.get_phone_msg_log()) #存放的手机信息 ll_list = getPhoneMsg(mc.get_phone_msg_log()) # monkey开始测试 sum = mc.get_sum() temp = "" monkeylog = "" start_monkey(mc.get_cmd(), mc.get_logdir(), mc.get_now(), mc.get_logcatname()) for i in range(sum): time.sleep(1) print(i) dn = dt.datetime.now() CPU[0].append(dn) m.top_cpu(mc.get_package_name()) MEN[0].append(dn) m.get_men(mc.get_package_name()) monkeylog = open(mc.get_logdir() + "\\" + mc.get_now()+"monkey.log") temp = monkeylog.read() monkeylog.close() if temp.count('Monkey finished')>0: print("测试完成咯") CPU[1].append(m.cpu) MEN[1].append(m.men) # geterror(ll_list, mc.get_log(), mc.get_remote_path(), mc.now) 错误显示 mt.cpu_men_plots(CPU, MEN) break else: print("设备不存在") else: print(u"配置文件不存在"+ini_file)
结果以曲线图展示
def cpu_men_plots(cpu, men): import matplotlib.pyplot as pl import matplotlib.dates as mdates import datetime # 处理异常数据,有时候得到数据(占用情况)会比时间多一次循环的数据,造成xy的数据不一致,而引起报错 if len(cpu[0]) != len(cpu[1][0]): cpu[1][0]= cpu[1][0][0:len(cpu[0])] if len(men[0]) != len(men[1][0]): men[1][0]= men[1][0][0:len(men[0])] print(men[0]) print(men[1][0]) a1 = pl.subplot(311) a1.set_title("CPU") a1.set_ylabel("占用情况%") a1.plot(cpu[0], cpu[1][0]) a1.xaxis.set_major_locator(mdates.SecondLocator(interval=1)) a1.xaxis.set_major_formatter(mdates.DateFormatter('%H:%M:%S')) a2 = pl.subplot(312) a2.set_title("内存") a2.set_ylabel("使用情况 K") a2.plot(men[0], men[1][0]) a2.xaxis.set_major_locator(mdates.SecondLocator(interval=2)) a2.xaxis.set_major_formatter(mdates.DateFormatter('%H:%M:%S')) # a3 = pl.subplot(313) # a3.set_title("流量") # a3.set_ylabel("使用情况 K") # a3.plot(x,list2) # a3.xaxis.set_major_formatter(mdates.DateFormatter('%H:%M:%S')) # a1.margins(x=0.2) pl.tight_layout() pl.show()
更多请参考我的源码
作者:望月成三人
链接:https://www.jianshu.com/p/f79493d431ff