linux 3d性能测试 arm,ARM平台常用性能测试方法

测试磁盘读写速度：

通过安装hdparm，可以测试EMMC的读写速度：

[email protected]:~$ sudo apt-get install hdparm [email protected]:/dev$ sudo hdparm -Tt /dev/mmcblk1 /dev/mmcblk1: Timing cached reads: 1202 MB in 2.00 seconds = 601.20 MB/sec Timing buffered disk reads: 340 MB in 3.01 seconds = 113.01 MB/sec [email protected]:~/data$ dd count=50 bs=1M if=/dev/zero of=~/data/test.img 50+0 records in 50+0 records out 52428800 bytes (52 MB, 50 MiB) copied, 0.620573 s, 84.5 MB/s

测试CPU性能

sysbench是一款开源的多线程性能测试工具，可以执行CPU/内存/线程/IO/数据库等方面的性能测试。可以通过apt install sysbench来安装。

CPU 测试时，会计算素数(对这个数字除以 2 到这个数字平方根之间的所有数字来验证素数)直到某个指定值所需要的时间。

sysbench --num-threads=4 --test=cpu --cpu-max-prime=20000 run

可以看到RK3399在该测试下的结果是9.9957s：

Threads started!

CPU speed:

events per second: 1927.70 General statistics: total time: 10.0027s total number of events: 19301 Latency (ms): min: 1.42 avg: 2.07 max: 22.63 95th percentile: 3.62 sum: 39982.94 Threads fairness: events (avg/stddev): 4825.2500/2066.85 execution time (avg/stddev): 9.9957/0.00

另一个性能测试跑分软件是nbench，可以对单个核进行内存、整型运算和浮点运算性能测试，具体包括以下10个测试项，可以通过wiki参考链接进一步了解。

Numeric sort - Sorts an array of long integers.

String sort - Sorts an array of strings of arbitrary length.

Bitfield - Executes a variety of bit manipulation functions.

Emulated floating-point - A small software floating-point package.

Fourier coefficients - A numerical analysis routine for calculating series approximations of waveforms.

Assignment algorithm - A well-known task allocation algorithm.

Huffman compression - A well-known text and graphics compression algorithm.

IDEA encryption - A relatively new block cipher algorithm.

Neural Net - A small but functional back-propagation network simulator.

LU Decomposition - A robust algorithm for solving linear equations.

获取源代码，并编译运行。

wget http://www.math.utah.edu/~mayer/linux/nbench-byte-2.2.3.tar.gz tar -xvzf nbench-byte-2.2.3.tar.gz cd nbench-byte-2.2.3 make ./nbench

可以看到6核的RK3399在nbench中跑分整数和浮点数分数79.409与44.893

[email protected]NEO4:~/nbench-byte-2.2.3$ ./nbench BYTEmark* Native Mode Benchmark ver. 2 (10/95) Index-split by Andrew D. Balsa (11/97) Linux/Unix* port by Uwe F. Mayer (12/96,11/97) TEST : Iterations/sec. : Old Index : New Index : : Pentium 90* : AMD K6/233* --------------------:------------------:-------------:------------ NUMERIC SORT : 1242.3 : 31.86 : 10.46 STRING SORT : 389.92 : 174.23 : 26.97 BITFIELD : 2.2521e+08 : 38.63 : 8.07 FP EMULATION : 369.93 : 177.51 : 40.96 FOURIER : 23017 : 26.18 : 14.70 ASSIGNMENT : 21.776 : 82.86 : 21.49 IDEA : 6673.9 : 102.07 : 30.31 HUFFMAN : 2230.3 : 61.85 : 19.75 NEURAL NET : 39.53 : 63.50 : 26.71 LU DECOMPOSITION : 1050.7 : 54.43 : 39.31 ==========================ORIGINAL BYTEMARK RESULTS========================== INTEGER INDEX : 79.409 FLOATING-POINT INDEX: 44.893 Baseline (MSDOS*) : Pentium* 90, 256 KB L2-cache, Watcom* compiler 10.0 ==============================LINUX DATA BELOW=============================== CPU : 6 CPU L2 Cache : OS : Linux 4.4.143 C compiler : gcc version 7.3.0 (Ubuntu/Linaro 7.3.0-27ubuntu1~18.04) libc : static MEMORY INDEX : 16.723 INTEGER INDEX : 22.505 FLOATING-POINT INDEX: 24.899 Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38 * Trademarks are property of their respective holder.

以下是S5P6818上使用32bit 系统的表现：

[email protected]Fire3:~/work/nbench-byte-2.2.3$ ./nbench BYTEmark* Native Mode Benchmark ver. 2 (10/95) Index-split by Andrew D. Balsa (11/97) Linux/Unix* port by Uwe F. Mayer (12/96,11/97) TEST : Iterations/sec. : Old Index : New Index : : Pentium 90* : AMD K6/233* --------------------:------------------:-------------:------------ NUMERIC SORT : 660.92 : 16.95 : 5.57 STRING SORT : 88.288 : 39.45 : 6.11 BITFIELD : 1.9179e+08 : 32.90 : 6.87 FP EMULATION : 102.92 : 49.38 : 11.40 FOURIER : 10112 : 11.50 : 6.46 ASSIGNMENT : 12.921 : 49.17 : 12.75 IDEA : 3181.4 : 48.66 : 14.45 HUFFMAN : 1202.3 : 33.34 : 10.65 NEURAL NET : 13.628 : 21.89 : 9.21 LU DECOMPOSITION : 459.08 : 23.78 : 17.17 ==========================ORIGINAL BYTEMARK RESULTS========================== INTEGER INDEX : 36.521 FLOATING-POINT INDEX: 18.158 Baseline (MSDOS*) : Pentium* 90, 256 KB L2-cache, Watcom* compiler 10.0 ==============================LINUX DATA BELOW=============================== CPU : 8 CPU L2 Cache : OS : Linux 4.4.49-s5p6818 C compiler : gcc version 5.4.0 20160609 (Ubuntu/Linaro 5.4.0-6ubuntu1~16.04.11) libc : libc-2.23.so MEMORY INDEX : 8.119 INTEGER INDEX : 9.939 FLOATING-POINT INDEX: 10.071 Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38 * Trademarks are property of their respective holder.

对应的相同平台上运行64bit系统的测试结果：

BYTEmark* Native Mode Benchmark ver. 2 (10/95) Index-split by Andrew D. Balsa (11/97) Linux/Unix* port by Uwe F. Mayer (12/96,11/97) TEST : Iterations/sec. : Old Index : New Index : : Pentium 90* : AMD K6/233* --------------------:------------------:-------------:------------ NUMERIC SORT : 800.68 : 20.53 : 6.74 STRING SORT : 159.48 : 71.26 : 11.03 BITFIELD : 2.22e+08 : 38.08 : 7.95 FP EMULATION : 219.6 : 105.37 : 24.32 FOURIER : 11728 : 13.34 : 7.49 ASSIGNMENT : 11.77 : 44.79 : 11.62 IDEA : 3420.5 : 52.32 : 15.53 HUFFMAN : 1133.7 : 31.44 : 10.04 NEURAL NET : 15.737 : 25.28 : 10.63 LU DECOMPOSITION : 532.56 : 27.59 : 19.92 ==========================ORIGINAL BYTEMARK RESULTS========================== INTEGER INDEX : 45.950 FLOATING-POINT INDEX: 21.031 Baseline (MSDOS*) : Pentium* 90, 256 KB L2-cache, Watcom* compiler 10.0 ==============================LINUX DATA BELOW=============================== CPU : 8 CPU L2 Cache : OS : Linux 4.4.49-s5p6818 C compiler : gcc version 7.3.0 (Ubuntu/Linaro 7.3.0-27ubuntu1~18.04) libc : libc-2.27.so MEMORY INDEX : 10.064 INTEGER INDEX : 12.645 FLOATING-POINT INDEX: 11.664 Baseline (LINUX) : AMD K6/233*, 512 KB L2-cache, gcc 2.7.2.3, libc-5.4.38 * Trademarks are property of their respective holder.

测试结果对比如下，运算性能上，整数提高了26%，浮点提高了16%：

测试项

32bit

64bit

性能差

NUMERIC SORT

660.92

800.68

21%

STRING SORT

88.288

159.48

81%

BITFIELD

1.92E+08

2.22E+08

16%

FP EMULATION

102.92

219.6

113%

FOURIER

10112

11728

16%

ASSIGNMENT

12.921

11.77

-9%

IDEA

3181.4

3420.5

8%

HUFFMAN

1202.3

1133.7

-6%

NEURAL NET

13.628

15.737

15%

LU DECOMPOSITION

459.08

532.56

16%

INTEGER

36.521

45.95

26%

FLOATING-POINT

18.158

21.031

16%

测试DDR性能

在小机上创建内存盘，将频繁读写的缓存和对速度有要求的小文件放到内存中，当然也可以用来测试读写速度：

[email protected]:/$ sudo mkdir /ram [email protected]:/$ sudo mount -t tmpfs -o size=100m,mode=0777 tmpfs /ram [email protected]:/$ cd ram [email protected]:/ram$ dd count=80 bs=1M if=/dev/zero of=/ram/test.img 50+0 records in 50+0 records out 52428800 bytes (52 MB, 50 MiB) copied, 0.134943 s, 389 MB/s

也可以用sysbench来测试，例如指定 4 个线程，缓冲区大小为 64KB，在内存中传输 10G 数据的测试：

sysbench --threads=4 --memory-block-size=64k --memory-total-size=10G memory run 10240.00 MiB transferred (3538.21 MiB/sec) General statistics: total time: 2.8907s total number of events: 163840 Latency (ms): min: 0.01 avg: 0.07 max: 11.02 95th percentile: 0.10 sum: 11207.68 Threads fairness: events (avg/stddev): 40960.0000/0.00 execution time (avg/stddev): 2.8019/0.02