Benchmarking some python crypto libraries

I wrote a little benchmark, comparing the PyCrypto and cryptography python modules.

You can find the script here.

results

Most symmetric ciphers are significantly faster in pyCrypto for small blocksizes. but cryptography is much faster for larger data.

The ratio is the byterate of pycrypto divided by the byterate of cryptography. So values less than 1 mean: cryptography is faster, values larger than 1: pycrypto is faster.

I found that the cryptography library is generally faster for large data volumes, while pycrypto is faster for encrypting small items.

symmetric crypto

Both pycrypto and cryptography have some call overhead. for pycrypto the byterate stabilizes for messages over 1K, while for `cryptography this happens for messages over 16K.

message size = 32 bytes, in Mbyte/sec

pycrypto cryptography ratio cipher
154 102 1.51 aes
129 91 1.41 blowfish
88 88 1.00 cast
52 71 0.74 des128
58 71 0.82 des192
102 73 1.41 des64
187 97 1.92 rc4

message size = 1048576 bytes, in Mbyte/sec

pycrypto cryptography ratio cipher
2054 22920 0.09 aes
326 936 0.35 blowfish
144 871 0.17 cast
79 231 0.34 des128
78 231 0.34 des192
212 234 0.90 des64
486 5050 0.10 rc4

hash algorithms

Both pycrypto and cryptography have some call overhead. for pycrypto the byterate stabilizes for messages over 16K, while for `cryptography this happens for messages over 1M.

message size = 32 bytes, in Mbyte/sec

pycrypto cryptography ratio algorithm
120 14 8.17 md5
114 14 7.65 ripemd
112 18 6.19 sha1
112 17 6.44 sha224
134 17 7.57 sha256
89 16 5.25 sha384
90 16 5.42 sha512

message size = 1048576 bytes, in Mbyte/sec

pycrypto cryptography ratio algorithm
5596 5231 1.07 md5
322 1660 0.19 ripemd
7871 6141 1.28 sha1
439 3498 0.13 sha224
390 3451 0.11 sha256
651 5219 0.12 sha384
611 5186 0.12 sha512

blocksize dependency

The Symmetric ciphers.

AES BLOWFISH CAST DES128 DES192 DES64 RC4 msgsize
1.51 1.41 1.00 0.74 0.82 1.41 1.92 32
1.30 1.12 0.65 0.59 0.63 1.36 1.23 64
1.21 0.85 0.38 0.48 0.49 1.23 0.88 128
0.99 0.68 0.30 0.38 0.44 1.22 0.47 256
0.78 0.62 0.25 0.38 0.38 1.16 0.35 512
0.56 0.51 -.-- 0.38 -.-- -.-- 0.22 1024
0.36 0.52 0.16 0.35 0.38 1.09 0.15 2048
0.24 0.44 0.17 0.33 0.39 1.00 0.13 4096
0.17 0.43 0.17 0.31 0.37 0.99 0.10 8192
0.12 0.39 0.17 0.32 0.36 1.03 0.10 16384
0.11 0.40 0.17 0.33 0.36 1.03 0.10 32768
0.09 0.36 0.17 0.37 0.35 1.06 0.08 65536
0.10 0.40 0.17 0.30 0.32 0.99 0.08 131072
0.09 0.37 0.15 0.33 0.35 0.85 0.07 262144
0.09 0.40 0.17 0.32 0.34 0.99 0.11 524288
0.09 0.35 0.17 0.34 0.34 0.90 0.10 1048576

And for hashing algorithms. For the SHAxxx algorithms cryptography is faster for messages over 1K. While for MD5 and SHA1 pyCrypto is always faster.

MD5 RIPEMD SHA1 SHA224 SHA256 SHA384 SHA512 msgsize
8.17 7.65 6.19 6.44 7.57 5.25 5.42 32
7.36 4.33 6.31 4.23 4.99 6.08 5.66 64
7.32 3.06 6.22 3.76 3.60 3.74 3.63 128
6.06 2.14 6.10 2.27 2.34 2.57 2.52 256
5.42 1.29 5.54 1.43 1.42 1.78 1.76 512
4.75 0.73 5.07 0.76 0.86 1.20 1.07 1024
3.17 0.45 3.67 0.51 0.48 0.72 0.65 2048
2.38 0.34 2.96 0.33 0.32 0.42 0.40 4096
1.79 0.30 2.26 0.20 0.23 0.29 0.27 8192
1.70 0.26 1.65 0.17 0.18 0.23 0.19 16384
1.32 0.24 1.31 0.14 0.16 0.16 0.16 32768
1.10 0.23 1.20 0.14 0.14 0.15 0.15 65536
1.05 0.23 1.05 0.13 0.14 0.13 0.13 131072
1.03 0.25 1.05 0.14 0.13 0.11 0.12 262144
1.04 0.21 1.00 0.13 0.13 0.12 0.12 524288
1.07 0.19 1.28 0.13 0.11 0.12 0.12 1048576

asymmetric crypto

in encryptions per second, all using a 512 bit message.

pow() pycrypto cryptography ratio modulus bits
17267 11827 34152 0.35 1024
4920 5014 17657 0.28 2048
1440 1717 6247 0.27 4096

So the cryptography library is generally faster. The pycrypto performance is roughly equal to using the pow() function.

random numbers

in Mbyte/sec

random sysrand pycrypto r/s s/p msgsize
478.0 55.9 2.0 8.6 27.4 32
726.4 70.9 3.1 10.2 22.6 64
1148.9 81.0 3.8 14.2 21.1 128
1476.4 88.1 4.4 16.8 20.1 256
1757.5 93.4 3.9 18.8 24.2 512
2141.6 94.3 2.6 22.7 36.6 1024
2339.3 100.2 1.9 23.3 53.4 2048
2405.1 94.8 1.1 25.4 83.1 4096
2361.0 101.9 0.6 23.2 158.7 8192
2568.1 97.7 0.3 26.3 286.8 16384

Conclusion: secure random numbers are expensive.

Comments

Comments powered by Disqus