[tahoe-dev] pycryptopp benchmarks on ARM CPU

[Zooko O'Whielacronx]

Okay, folks, pycryptopp is now working on one of our ARM CPU
buildbots, and I've run some measurements of the time to do different
crypto operations.

This machine has an ARM926EJ-S rev 0 (v5l) "Feroceon" at 266 MHz, with
32 KiB icache and 32 KiB dcache.

Here are a few crypto algorithm measurements:

Algorithm    MiB/Sec Cycle/Byte
---------    ------- ----------
MD5               24       10.7
SHA-1              9       28.5
SHA-256            6       42.5

Salsa20/12        13       19.0
AES/CTR (128-bit)  2      131.9

(By the way, the SHA-3 project looks like it will produce a hash
algorithm at least as fast as SHA-256 on this sort of CPU.  The
candidate which has the best performance on such a CPU so far is
"edonr256" which takes 17.5 cycles/byte on a similar ARM CPU:
http://bench.cr.yp.to/results-hash.html#arm-apollo .)

Algorithm          millisec/Op megacycle/Op
---------          ----------- ------------
RSA 2056 Sig            950.00       252.70
RSA 2056 Verify          12.82         3.41
ECDSA secp256k1 Sig      49.05        13.05
ECDSA secp256k1 Sig/pre  31.56         8.40
ECDSA secp256k1 Ver     128.75        34.25
ECDSA secp256k1 Ver/pre  51.00        13.57
ECDSA secp192r1 Sig      32.90         8.75
ECDSA secp192r1 Sig/pre  16.95         4.51
ECDSA secp192r1 Ver      58.82        15.65
ECDSA secp192r1 Ver/pre  28.86         7.68

And here is the big issue -- the main reason why we're switching to
ECDSA in the next release of Tahoe-LAFS (not the imminent v1.5
release).  It is the time it takes to generate a new key pair.  (I
suspect that the RSA keypair generation in Crypto++ is not optimized.)

Algorithm       millisec/new key
---------       ---------------
RSA 2056                31,670    (best)
RSA 2056                57,860    (mean)
RSA 2056                98,180    (worst)
ECDSA secp192r1              2.75 (best)
ECDSA secp192r1              2.79 (mean)
ECDSA secp192r1              2.89 (worst)