9468305/Android md5 blake2 性能优化测试数据.txt

## Android md5 blake2 性能优化测试数据.txt
Android手机测试数据

测试文件：460MB

源码：见 https://gist.github.com/9468305/97dca7c470ee02a6867c

使用场景：需求源于这里 https://gist.github.com/9468305/fa8f1307ea4738225fca

测试思路：测试mmap，file io buffer read，OpenMP并行数，对各种数据摘要算法，在不同手机上的性能表现进行统计分析

首先使用mmap映射整个文件到内存，直接使用460MB内存地址计算
blake2sp 和 blake2bp 并行数=官网默认值8

华为荣耀3X ndk-build thumb2 gcc -O3 执行2次
blake2s     time = 15.230977 seconds    15.149100 seconds
blake2b     time = 23.496808 seconds    23.273819 seconds
blake2sp    time = 3.458435 seconds     3.083542 seconds
blake2bp    time = 6.946069 seconds     7.100667 seconds


华为荣耀3X ndk-build thumb2 gcc -O2
run test on device
blake2s     time = 15.519730 seconds
blake2b     time = 22.998304 seconds
blake2sp    time = 3.352492 seconds
blake2bp    time = 6.810990 seconds


华为荣耀3X ndk-build thumb2 gcc -Os
blake2s     time = 15.541925 seconds
blake2b     time = 22.988108 seconds
blake2sp    time = 3.389670 seconds
blake2bp    time = 6.830233 seconds


华为荣耀3X ndk-build arm gcc -O3
blake2s     time = 15.135739 seconds
blake2b     time = 23.559449 seconds
blake2sp    time = 3.393177 seconds
blake2bp    time = 6.834560 seconds


数据分析：thumb arm相同，-O2 -O3 -O3 相同；
继续测试：以下使用thumb -O3


小米4 -O3 thumb 均衡模式（运行期间可能手机黑屏了，导致blake2bp数据不准）
blake2s     time = 11.874335 seconds
blake2b     time = 31.524835 seconds
blake2sp    time = 3.485516 seconds
blake2bp    time = 20.172217 seconds


小米4 省电模式(CPU被降频)
blake2s     time = 21.074071 seconds
blake2b     time = 48.200784 seconds
blake2sp    time = 9.167828 seconds
blake2bp    time = 23.293334 seconds


小米4 性能模式(CPU频率不锁，可动态至最高，但不是始终最高频运行)
blake2s     time = 6.652548 seconds
blake2b     time = 20.712190 seconds
blake2sp    time = 2.550314 seconds
blake2bp    time = 6.108460 seconds


Nexus5 thumb2 16指令集 执行2次 （Nexus5比较奇葩，32位CPU+64位OS，表现各种异常）
blake2s     time = 6.078702 seconds     6.857189 seconds
blake2b     time = 19.730120 seconds    19.689735 seconds
blake2sp    time = 8.525794 seconds     10.762844 seconds
blake2bp    time = 21.908520 seconds    23.456138 seconds


Nexus5 arm 32位指令集
blake2s     time = 6.817943 seconds
blake2b     time = 20.754325 seconds
blake2sp    time = 10.346421 seconds
blake2bp    time = 23.718929 seconds


数据分析：并行数可能对执行效率有影响
继续测试：修改blake2sp blake2bp的并行数=4

Nexus5 执行2次 （可能中途手机进入省电模式）
blake2s     time = 10.197930 seconds    10.092280 seconds
blake2b     time = 15.541306 seconds    15.102294 seconds
blake2sp    time = 3.458573 seconds     3.345040 seconds
blake2bp    time = 4.747609 seconds     4.683213 seconds


小米4 执行2次
blake2s     time = 6.572716 seconds     6.555986 seconds
blake2b     time = 20.772833 seconds    20.788557 seconds
blake2sp    time = 2.553117 seconds     2.616197 seconds
blake2bp    time = 6.034184 seconds     6.170549 seconds


Nexus5 执行2次 （手机亮屏）
blake2s     time = 5.844344 seconds     5.960309 seconds
blake2b     time = 19.377594 seconds    19.664046 seconds
blake2sp    time = 7.031088 seconds     7.153667 seconds
blake2bp    time = 21.665652 seconds    21.794273 seconds


数据分析：blake2b blake2bp 在32位OS上性能太差; 并行数对测试结果有影响；
继续测试：放弃blake2b，blake2bp，观察blake2s blake2sp在并行数=2的表现；


小米4
blake2s     time = 6.574686 seconds
blake2sp    time = 3.746227 seconds


华为荣耀3X
blake2s     time = 10.452700 seconds
blake2sp    time = 6.534018 seconds


Nexus5
blake2s     time = 6.045295 seconds
blake2sp    time = 6.597525 seconds


测试分析：对于小米4 华为荣耀3X，并行数2，4，8的差异不大；对于Nexus5，并行数越小，性能越高，但弱于其他手机（可是这款手机的CPU不算差）
继续测试：改用c file io读取文件，buffer = 8KB；不使用mmap；

华为荣耀3X 执行2次
blake2s     time = 10.288910 seconds    10.249259 seconds
blake2sp    time = 4.657944 seconds     4.876143 seconds


Nexus5 执行2次
blake2s     time = 5.689329 seconds     6.569514 seconds
blake2sp    time = 38.843955 seconds    36.973709 seconds


小米4
blake2s     time = 7.024298 seconds     6.848578 seconds
blake2sp    time = 5.210849 seconds     5.023147 seconds


小米4 io buffer = 16KB 结果跟8KB没差别
blake2s     time = 6.938758 seconds     6.939108 seconds
blake2sp    time = 11.459279 seconds    22.057534 seconds 手机中途黑屏 CPU降频


小米4 io buffer = 1Byte blake2sp代码中没有对入参size做判断，对1Byte数据做并行计算发生异常
blake2s     time = 24.402502 seconds
blake2sp    fail - 死循环


小米4 io buffer = 32KB
blake2s     time = 6.626112 seconds 6.779171 seconds 跟8KB没差别
blake2sp    time = 4.760458 seconds 4.642409 seconds 差距出现了


数据分析：io buffer大小对串行计算影响不大；对并行计算影响很大；
        阅读blake2sp源码得知，它对每次update的buffer做并行计算，因此buffer越小，创建线程的开销越高，导致运行更慢。
        而mmap整个文件进行read，规避了这个代码缺陷
继续测试：使用新的并行方案，将整个文件按并行数切分，每段数据使用file io buffer读写
        blake2sp_file 源码见：https://gist.github.com/9468305/20068a1af16910361278


小米4 执行4次
blake2s         time = 6.881021 seconds 6.881021 seconds 6.890889 seconds 6.816782 seconds
blake2sp        time = 4.752358 seconds 4.752358 seconds 6.570885 seconds 4.864570 seconds
blake2sp_file   time = 2.652560 seconds 2.652560 seconds 2.652320 seconds 2.791247 seconds
增加md5标准方案（串行）
md5             time = 4.476925 seconds time = 4.487960 seconds


华为荣耀3X 执行2次
blake2s         time = 11.003261 seconds    11.111040 seconds
blake2sp        time = 7.405163 seconds     6.993986 seconds
blake2sp_file   time = 3.407042 seconds     4.195505 seconds
md5             time = 3.657673 seconds     3.724206 seconds


Nexus5
blake2s         time = 5.720441 seconds
blake2sp        time = 30.024116 seconds
blake2sp_file   time = 7.610792 seconds
md5             time = 3.173777 seconds


数据分析：md5串行优于blake2各种变种方案；blake2sp_file优于blake2sp
继续测试：添加md5标准串行和md5文件并行方案，观察性能; 首先md5并行数=8
        增加MD5并行方案md5p 源码见 https://gist.github.com/9468305/97dca7c470ee02a6867c


华为荣耀3X
blake2s         time = 10.355605 seconds    10.297682 seconds
blake2sp        time = 4.699469 seconds     4.719653 seconds
blake2sp_file   time = 3.587286 seconds     3.332196 seconds
md5             time = 3.927048 seconds     3.923544 seconds
md5p            time = 1.160357 seconds     1.052565 seconds


Nexus5
blake2s         time = 5.444591 seconds     5.516199 seconds
blake2sp        time = 34.665423 seconds    36.844409 seconds
blake2sp_file   time = 7.044548 seconds     7.199510 seconds
md5             time = 2.865387 seconds     3.132674 seconds
md5p            time = 3.096923 seconds     3.492756 seconds


小米2S 均衡模式
blake2s         time = 10.639981 seconds
blake2sp        time = 11.367557 seconds
blake2sp_file   time = 4.067023 seconds
md5             time = 7.117229 seconds
md5p            time = 3.617916 seconds

小米2S 性能模式
blake2s         time = 11.341095 seconds
blake2sp        time = 7.885427 seconds
blake2sp_file   time = 3.922051 seconds
md5             time = 7.437327 seconds
md5p            time = 3.234458 seconds


华为荣耀3X
blake2s         time = 11.012500 seconds
blake2sp        time = 6.982329 seconds
blake2sp_file   time = 3.445275 seconds
md5             time = 3.889710 seconds
md5p            time = 1.171731 seconds


华为荣耀3X md5并行数=8
blake2s         time = 10.212562 seconds
blake2sp        time = 4.839898 seconds
blake2sp_file   time = 3.537396 seconds
md5             time = 3.851837 seconds
md5p            time = 0.819049 seconds


华为荣耀3X 所有算法的并行数=8
blake2s         time = 10.210714 seconds
blake2sp        time = 5.407414 seconds
blake2sp_file   time = 1.791010 seconds
md5             time = 3.841578 seconds
md5p            time = 0.606765 seconds


小米4 并行数=8
blake2s         time = 7.481798 seconds     6.739367 seconds
blake2sp        time = 10.937392 seconds    10.036826 seconds
blake2sp_file   time = 2.343030 seconds     2.327568 seconds
md5             time = 4.670945 seconds     4.498684 seconds
md5p            time = 1.817163 seconds     1.622922 seconds


小米4 并行数=4
blake2s         time = 6.784404 seconds     6.810516 seconds
blake2sp        time = 10.230556 seconds    9.958661 seconds
blake2sp_file   time = 2.353576 seconds     2.342131 seconds
md5             time = 4.474723 seconds     4.493030 seconds
md5p            time = 1.787866 seconds     1.813598 seconds


Nexus5 md5并行数=4
blake2s         time = 5.591858 seconds     5.504601 seconds
blake2sp        time = 20.066783 seconds    time = 20.679386 seconds
blake2sp_file   time = 7.147342 seconds     7.157790 seconds
md5             time = 2.969569 seconds     2.886641 seconds
md5p            time = 3.274280 seconds     3.064426 seconds


Nexus5 md5并行数=8
blake2s         time = 6.177822 seconds
blake2sp        time = 19.658719 seconds
blake2sp_file   time = 7.498131 seconds
md5             time = 3.202704 seconds
md5p            time = 3.490773 seconds


数据分析：blake2sp_file优于blake2sp；md5p优于md5；file分段read方案时，并行数已经影响不大
继续测试：淘汰blake2sp


Nexus5
blake2s         time = 5.575001 seconds
blake2sp_file   time = 7.127042 seconds
md5             time = 2.874951 seconds
md5p            time = 3.333800 seconds


华为荣耀3X md5p=4
blake2s         time = 11.114438 seconds    10.220039 seconds
blake2sp_file   time = 2.727083 seconds     1.895031 seconds
md5             time = 3.851460 seconds     3.850437 seconds
md5p            time = 1.082791 seconds     0.995741 seconds


华为荣耀3X md5p=8
blake2s         time = 10.171139 seconds    10.433916 seconds
blake2sp_file   time = 1.989670 seconds     2.010275 seconds
md5             time = 3.873485 seconds     3.869215 seconds
md5p            time = 0.618082 seconds     0.664903 seconds


小米4 md5p=8
blake2s         time = 6.755866 seconds     6.765623 seconds
blake2sp_file   time = 2.260821 seconds     2.245780 seconds
md5             time = 4.471778 seconds     4.477018 seconds
md5p            time = 1.802037 seconds     1.611127 seconds

总结分析：
Nexus5奇葩，总是不按常理出牌；
在各种情况下，md5p表现都优秀；
并行数=4在各种情况下表现稳定；
华为荣耀3X应该是8核手机，所以并行数=8时总是最优；
目前主流手机是4核为主，因此选择并行数=4的md5算法最好；

后续测试：
进行大文件的分段mmap+并行算法进行测试

存在争议：
究竟瓶颈是cpu还是file io？
	Android手机测试数据

	测试文件：460MB

	源码：见 https://gist.github.com/9468305/97dca7c470ee02a6867c

	使用场景：需求源于这里 https://gist.github.com/9468305/fa8f1307ea4738225fca

	测试思路：测试mmap，file io buffer read，OpenMP并行数，对各种数据摘要算法，在不同手机上的性能表现进行统计分析

	首先使用mmap映射整个文件到内存，直接使用460MB内存地址计算
	blake2sp 和 blake2bp 并行数=官网默认值8

	华为荣耀3X ndk-build thumb2 gcc -O3 执行2次
	blake2s time = 15.230977 seconds 15.149100 seconds
	blake2b time = 23.496808 seconds 23.273819 seconds
	blake2sp time = 3.458435 seconds 3.083542 seconds
	blake2bp time = 6.946069 seconds 7.100667 seconds


	华为荣耀3X ndk-build thumb2 gcc -O2
	run test on device
	blake2s time = 15.519730 seconds
	blake2b time = 22.998304 seconds
	blake2sp time = 3.352492 seconds
	blake2bp time = 6.810990 seconds


	华为荣耀3X ndk-build thumb2 gcc -Os
	blake2s time = 15.541925 seconds
	blake2b time = 22.988108 seconds
	blake2sp time = 3.389670 seconds
	blake2bp time = 6.830233 seconds


	华为荣耀3X ndk-build arm gcc -O3
	blake2s time = 15.135739 seconds
	blake2b time = 23.559449 seconds
	blake2sp time = 3.393177 seconds
	blake2bp time = 6.834560 seconds


	数据分析：thumb arm相同，-O2 -O3 -O3 相同；
	继续测试：以下使用thumb -O3


	小米4 -O3 thumb 均衡模式（运行期间可能手机黑屏了，导致blake2bp数据不准）
	blake2s time = 11.874335 seconds
	blake2b time = 31.524835 seconds
	blake2sp time = 3.485516 seconds
	blake2bp time = 20.172217 seconds


	小米4 省电模式(CPU被降频)
	blake2s time = 21.074071 seconds
	blake2b time = 48.200784 seconds
	blake2sp time = 9.167828 seconds
	blake2bp time = 23.293334 seconds


	小米4 性能模式(CPU频率不锁，可动态至最高，但不是始终最高频运行)
	blake2s time = 6.652548 seconds
	blake2b time = 20.712190 seconds
	blake2sp time = 2.550314 seconds
	blake2bp time = 6.108460 seconds


	Nexus5 thumb2 16指令集执行2次（Nexus5比较奇葩，32位CPU+64位OS，表现各种异常）
	blake2s time = 6.078702 seconds 6.857189 seconds
	blake2b time = 19.730120 seconds 19.689735 seconds
	blake2sp time = 8.525794 seconds 10.762844 seconds
	blake2bp time = 21.908520 seconds 23.456138 seconds


	Nexus5 arm 32位指令集
	blake2s time = 6.817943 seconds
	blake2b time = 20.754325 seconds
	blake2sp time = 10.346421 seconds
	blake2bp time = 23.718929 seconds


	数据分析：并行数可能对执行效率有影响
	继续测试：修改blake2sp blake2bp的并行数=4

	Nexus5 执行2次（可能中途手机进入省电模式）
	blake2s time = 10.197930 seconds 10.092280 seconds
	blake2b time = 15.541306 seconds 15.102294 seconds
	blake2sp time = 3.458573 seconds 3.345040 seconds
	blake2bp time = 4.747609 seconds 4.683213 seconds


	小米4 执行2次
	blake2s time = 6.572716 seconds 6.555986 seconds
	blake2b time = 20.772833 seconds 20.788557 seconds
	blake2sp time = 2.553117 seconds 2.616197 seconds
	blake2bp time = 6.034184 seconds 6.170549 seconds


	Nexus5 执行2次（手机亮屏）
	blake2s time = 5.844344 seconds 5.960309 seconds
	blake2b time = 19.377594 seconds 19.664046 seconds
	blake2sp time = 7.031088 seconds 7.153667 seconds
	blake2bp time = 21.665652 seconds 21.794273 seconds


	数据分析：blake2b blake2bp 在32位OS上性能太差; 并行数对测试结果有影响；
	继续测试：放弃blake2b，blake2bp，观察blake2s blake2sp在并行数=2的表现；


	小米4
	blake2s time = 6.574686 seconds
	blake2sp time = 3.746227 seconds


	华为荣耀3X
	blake2s time = 10.452700 seconds
	blake2sp time = 6.534018 seconds


	Nexus5
	blake2s time = 6.045295 seconds
	blake2sp time = 6.597525 seconds


	测试分析：对于小米4 华为荣耀3X，并行数2，4，8的差异不大；对于Nexus5，并行数越小，性能越高，但弱于其他手机（可是这款手机的CPU不算差）
	继续测试：改用c file io读取文件，buffer = 8KB；不使用mmap；

	华为荣耀3X 执行2次
	blake2s time = 10.288910 seconds 10.249259 seconds
	blake2sp time = 4.657944 seconds 4.876143 seconds


	Nexus5 执行2次
	blake2s time = 5.689329 seconds 6.569514 seconds
	blake2sp time = 38.843955 seconds 36.973709 seconds


	小米4
	blake2s time = 7.024298 seconds 6.848578 seconds
	blake2sp time = 5.210849 seconds 5.023147 seconds


	小米4 io buffer = 16KB 结果跟8KB没差别
	blake2s time = 6.938758 seconds 6.939108 seconds
	blake2sp time = 11.459279 seconds 22.057534 seconds 手机中途黑屏 CPU降频


	小米4 io buffer = 1Byte blake2sp代码中没有对入参size做判断，对1Byte数据做并行计算发生异常
	blake2s time = 24.402502 seconds
	blake2sp fail - 死循环


	小米4 io buffer = 32KB
	blake2s time = 6.626112 seconds 6.779171 seconds 跟8KB没差别
	blake2sp time = 4.760458 seconds 4.642409 seconds 差距出现了


	数据分析：io buffer大小对串行计算影响不大；对并行计算影响很大；
	阅读blake2sp源码得知，它对每次update的buffer做并行计算，因此buffer越小，创建线程的开销越高，导致运行更慢。
	而mmap整个文件进行read，规避了这个代码缺陷
	继续测试：使用新的并行方案，将整个文件按并行数切分，每段数据使用file io buffer读写
	blake2sp_file 源码见：https://gist.github.com/9468305/20068a1af16910361278


	小米4 执行4次
	blake2s time = 6.881021 seconds 6.881021 seconds 6.890889 seconds 6.816782 seconds
	blake2sp time = 4.752358 seconds 4.752358 seconds 6.570885 seconds 4.864570 seconds
	blake2sp_file time = 2.652560 seconds 2.652560 seconds 2.652320 seconds 2.791247 seconds
	增加md5标准方案（串行）
	md5 time = 4.476925 seconds time = 4.487960 seconds


	华为荣耀3X 执行2次
	blake2s time = 11.003261 seconds 11.111040 seconds
	blake2sp time = 7.405163 seconds 6.993986 seconds
	blake2sp_file time = 3.407042 seconds 4.195505 seconds
	md5 time = 3.657673 seconds 3.724206 seconds


	Nexus5
	blake2s time = 5.720441 seconds
	blake2sp time = 30.024116 seconds
	blake2sp_file time = 7.610792 seconds
	md5 time = 3.173777 seconds


	数据分析：md5串行优于blake2各种变种方案；blake2sp_file优于blake2sp
	继续测试：添加md5标准串行和md5文件并行方案，观察性能; 首先md5并行数=8
	增加MD5并行方案md5p 源码见 https://gist.github.com/9468305/97dca7c470ee02a6867c


	华为荣耀3X
	blake2s time = 10.355605 seconds 10.297682 seconds
	blake2sp time = 4.699469 seconds 4.719653 seconds
	blake2sp_file time = 3.587286 seconds 3.332196 seconds
	md5 time = 3.927048 seconds 3.923544 seconds
	md5p time = 1.160357 seconds 1.052565 seconds


	Nexus5
	blake2s time = 5.444591 seconds 5.516199 seconds
	blake2sp time = 34.665423 seconds 36.844409 seconds
	blake2sp_file time = 7.044548 seconds 7.199510 seconds
	md5 time = 2.865387 seconds 3.132674 seconds
	md5p time = 3.096923 seconds 3.492756 seconds


	小米2S 均衡模式
	blake2s time = 10.639981 seconds
	blake2sp time = 11.367557 seconds
	blake2sp_file time = 4.067023 seconds
	md5 time = 7.117229 seconds
	md5p time = 3.617916 seconds

	小米2S 性能模式
	blake2s time = 11.341095 seconds
	blake2sp time = 7.885427 seconds
	blake2sp_file time = 3.922051 seconds
	md5 time = 7.437327 seconds
	md5p time = 3.234458 seconds


	华为荣耀3X
	blake2s time = 11.012500 seconds
	blake2sp time = 6.982329 seconds
	blake2sp_file time = 3.445275 seconds
	md5 time = 3.889710 seconds
	md5p time = 1.171731 seconds


	华为荣耀3X md5并行数=8
	blake2s time = 10.212562 seconds
	blake2sp time = 4.839898 seconds
	blake2sp_file time = 3.537396 seconds
	md5 time = 3.851837 seconds
	md5p time = 0.819049 seconds


	华为荣耀3X 所有算法的并行数=8
	blake2s time = 10.210714 seconds
	blake2sp time = 5.407414 seconds
	blake2sp_file time = 1.791010 seconds
	md5 time = 3.841578 seconds
	md5p time = 0.606765 seconds


	小米4 并行数=8
	blake2s time = 7.481798 seconds 6.739367 seconds
	blake2sp time = 10.937392 seconds 10.036826 seconds
	blake2sp_file time = 2.343030 seconds 2.327568 seconds
	md5 time = 4.670945 seconds 4.498684 seconds
	md5p time = 1.817163 seconds 1.622922 seconds


	小米4 并行数=4
	blake2s time = 6.784404 seconds 6.810516 seconds
	blake2sp time = 10.230556 seconds 9.958661 seconds
	blake2sp_file time = 2.353576 seconds 2.342131 seconds
	md5 time = 4.474723 seconds 4.493030 seconds
	md5p time = 1.787866 seconds 1.813598 seconds


	Nexus5 md5并行数=4
	blake2s time = 5.591858 seconds 5.504601 seconds
	blake2sp time = 20.066783 seconds time = 20.679386 seconds
	blake2sp_file time = 7.147342 seconds 7.157790 seconds
	md5 time = 2.969569 seconds 2.886641 seconds
	md5p time = 3.274280 seconds 3.064426 seconds


	Nexus5 md5并行数=8
	blake2s time = 6.177822 seconds
	blake2sp time = 19.658719 seconds
	blake2sp_file time = 7.498131 seconds
	md5 time = 3.202704 seconds
	md5p time = 3.490773 seconds


	数据分析：blake2sp_file优于blake2sp；md5p优于md5；file分段read方案时，并行数已经影响不大
	继续测试：淘汰blake2sp


	Nexus5
	blake2s time = 5.575001 seconds
	blake2sp_file time = 7.127042 seconds
	md5 time = 2.874951 seconds
	md5p time = 3.333800 seconds


	华为荣耀3X md5p=4
	blake2s time = 11.114438 seconds 10.220039 seconds
	blake2sp_file time = 2.727083 seconds 1.895031 seconds
	md5 time = 3.851460 seconds 3.850437 seconds
	md5p time = 1.082791 seconds 0.995741 seconds


	华为荣耀3X md5p=8
	blake2s time = 10.171139 seconds 10.433916 seconds
	blake2sp_file time = 1.989670 seconds 2.010275 seconds
	md5 time = 3.873485 seconds 3.869215 seconds
	md5p time = 0.618082 seconds 0.664903 seconds


	小米4 md5p=8
	blake2s time = 6.755866 seconds 6.765623 seconds
	blake2sp_file time = 2.260821 seconds 2.245780 seconds
	md5 time = 4.471778 seconds 4.477018 seconds
	md5p time = 1.802037 seconds 1.611127 seconds

	总结分析：
	Nexus5奇葩，总是不按常理出牌；
	在各种情况下，md5p表现都优秀；
	并行数=4在各种情况下表现稳定；
	华为荣耀3X应该是8核手机，所以并行数=8时总是最优；
	目前主流手机是4核为主，因此选择并行数=4的md5算法最好；

	后续测试：
	进行大文件的分段mmap+并行算法进行测试

	存在争议：
	究竟瓶颈是cpu还是file io？