Add benchmarks for generation and fingerprinting

[dmcgowan]

Results with i7-3687U CPU @ 2.10GHz × 4

BenchmarkECP256Generate     5000        765631 ns/op
BenchmarkECP384Generate      200      10228670 ns/op
BenchmarkECP521Generate      100      17862164 ns/op
BenchmarkRSA2048Generate           2     911764954 ns/op
BenchmarkRSA3072Generate           1    1426658951 ns/op
BenchmarkRSA4096Generate           1    19010220221 ns/op
BenchmarkECP256Fingerprint    500000          3202 ns/op
BenchmarkECP384Fingerprint    500000          3304 ns/op
BenchmarkECP521Fingerprint    500000          4004 ns/op
BenchmarkRSA2048Fingerprint       10     174062887 ns/op
BenchmarkRSA3072Fingerprint        1    6930198558 ns/op
BenchmarkRSA4096Fingerprint        1    3772528822 ns/op

[dmcgowan]

@jlhawn was the slow RSA fingerprint generation the reason we didn't use the standard encoding before? I believe the RSA encoding does some prime verification, perhaps if we needed speed for RSA we could take that code and strip out the prime check.

[jlhawn]

your fingerprint benchmark is counting the key generation. I've fixed it here: jlhawn/libtrust@ac2e85a

Here are my results:

BenchmarkECP256Generate         5000         617555 ns/op   ~0.6 ms
BenchmarkECP384Generate          200        8964434 ns/op   ~9.0 ms
BenchmarkECP521Generate          100       15907002 ns/op  ~16.0 ms
BenchmarkRSA2048Generate           5      474526423 ns/op   ~475 ms
BenchmarkRSA3072Generate           1     1709847534 ns/op   ~1.71 s
BenchmarkRSA4096Generate           1    17560242292 ns/op   ~17.6 s
BenchmarkECP256Fingerprint   1000000           3059 ns/op
BenchmarkECP384Fingerprint   1000000           3090 ns/op
BenchmarkECP521Fingerprint    500000           3528 ns/op
BenchmarkRSA2048Fingerprint   500000           4867 ns/op
BenchmarkRSA3072Fingerprint   500000           5725 ns/op
BenchmarkRSA4096Fingerprint   500000           6061 ns/op

[jlhawn]

The RSA key generation is slow due to checking for a large prime and it gets exponentially more difficult (half a second for a 2048-bit RSA key vs 17.5 seconds for a 4096-bit RSA key). I also recall a TLS config benchmark that showed loading an RSA cert/key as slow because the go standard lib unnecessarily decides to do a prime check again when loading in the private key.

[jlhawn]

Also, the RSA key generation is probabilistic. It doesn't always take 17 seconds to generate... You may get lucky and guess a prime number early on in the search, you may be unlucky and have it take over a minute! I'm not sure what it would be on expectation because the go bench only ever runs one iteration, and the variance seems pretty high.

[dmcgowan]

@jlhawn Good catch on the benchmark, incorporated your change and merging