GMP vs BN

[Torbjorn Granlund]

mts.spb.suxx at mail.ru writes:

    Did anyone try to benchmark the GMP against the BigNum library from
    the OpenSSL package?
  
Here are some results.  I used the command "openssl speed rsa" to
get the OpenSSL numbers and a small home frown program for RSA to
get the GMP numbers.

Athlon 2.2 GHz  (32-bit):
  OpenSSL 0.9.7d    sign    verify    sign/s verify/s
  rsa  512 bits   0.0006s   0.0001s   1668.5  19312.4
  rsa 1024 bits   0.0024s   0.0001s    410.8   8022.4
  rsa 2048 bits   0.0130s   0.0004s     77.1   2690.4
  rsa 4096 bits   0.0815s   0.0013s     12.3    797.9
  GMP 4.1.4         sign    verify    sign/s verify/s
  rsa  512 bits   0.00023s  0.00003s  4262.3  39428.0
  rsa 1024 bits   0.00131s  0.00007s   764.2  14287.4
  rsa 2048 bits   0.00822s  0.00022s   121.6   4448.7
  rsa 4096 bits   0.05564s  0.00070s    18.0   1435.8
GMP is 2 to 2.5 times faster here.

Athlon 1.8 GHz  (64-bit):
  OpenSSL 0.9.7d    sign    verify    sign/s verify/s
  rsa  512 bits   0.0003s   0.0000s   3194.6  36893.6
  rsa 1024 bits   0.0009s   0.0001s   1135.5  17194.5
  rsa 2048 bits   0.0047s   0.0002s    213.9   6609.9
  rsa 4096 bits   0.0288s   0.0005s     34.7   2158.1
  GMP 4.1.4         sign    verify    sign/s verify/s
  rsa  512 bits   0.00016s  0.00001s  6225.3  67309.0
  rsa 1024 bits   0.00081s  0.00004s  1230.3  24074.8
  rsa 2048 bits   0.00494s  0.00014s   202.5   7206.2
  rsa 4096 bits   0.03373s  0.00049s    29.6   2037.2
Here OpenSSL actually beats GMP for 3 of the 8 tests.  (Since GMP
lacks assembly loops for this platform, the compiler's performance
for the critical loops is crucial.  I used GCC 3.4.2 with the
odd options -O2 -funroll-loops -mtune=nocona.

Pentium 4 (Northwood) 2.6 GHz:
  OpenSSL 0.9.7d    sign    verify    sign/s verify/s
  rsa  512 bits   0.0009s   0.0001s   1147.5  12017.8
  rsa 1024 bits   0.0042s   0.0002s    236.7   4317.9
  rsa 2048 bits   0.0253s   0.0008s     39.5   1327.2
  rsa 4096 bits   0.1689s   0.0026s      5.9    378.3
  GMP 4.1.4         sign    verify    sign/s verify/s
  rsa  512 bits   0.00028s  0.00003s  3529.2  29529.7
  rsa 1024 bits   0.00164s  0.00009s   611.3  10630.4
  rsa 2048 bits   0.01055s  0.00030s    94.8   3360.0
  rsa 4096 bits   0.06883s  0.00096s    14.5   1042.5
Here the order is restored, GMP is around 3 times faster than
OpenSSL.

Alpha 21264 833 MHz:
  OpenSSL 0.9.7e    sign    verify    sign/s verify/s
  rsa  512 bits   0.0007s   0.0001s   1416.0  15786.1
  rsa 1024 bits   0.0022s   0.0001s    456.4   7431.2
  rsa 2048 bits   0.0117s   0.0004s     85.3   2726.4
  rsa 4096 bits   0.0725s   0.0012s     13.8    835.5
  GMP 4.1.4         sign    verify    sign/s verify/s
  rsa  512 bits   0.00031s  0.00003s  3185.7  33465.3
  rsa 1024 bits   0.00141s  0.00007s   707.6  13452.9
  rsa 2048 bits   0.00772s  0.00021s   129.5   4809.2
  rsa 4096 bits   0.04662s  0.00065s    21.4   1548.2
Again, GMP outperforms OpenSSL.  On average about twice
as fast.

This can continue with many platforms.  On OpenSSL's home turf
(RSA), GMP 4.1.4 is usually between 2 and 3 times faster.

That factor will substantially increase in the future.  :-)

--
Torbjörn