Calibrate with fixed r

[LouisChrist]

[edit by @elithrar]: Addresses #12

[elithrar]

@FiloSottile - mind taking a look at this as well? More eyes always good.

[elithrar]

Failed in Go 1.7 - restarted the build.

$ go test -v ./...
=== RUN   TestGenerateRandomBytes
--- PASS: TestGenerateRandomBytes (0.00s)
=== RUN   TestGenerateFromPassword
--- PASS: TestGenerateFromPassword (9.47s)
=== RUN   TestCompareHashAndPassword
--- PASS: TestCompareHashAndPassword (0.18s)
=== RUN   TestCost
--- PASS: TestCost (0.06s)
=== RUN   TestDecodeHash
--- PASS: TestDecodeHash (0.00s)
=== RUN   TestCalibrate
--- FAIL: TestCalibrate (10.40s)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:131072, R:8, P:1, SaltLen:16, DKLen:32} took 472.320073ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:131072, R:8, P:1, SaltLen:16, DKLen:32} took 506.033718ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:65536, R:8, P:1, SaltLen:16, DKLen:32} took 230.961435ms (<nil>)
	scrypt_test.go:145: 2. GenerateFromPassword was too fast (wanted around 500ms, got 230.961435ms) with scrypt.Params{N:65536, R:8, P:1, SaltLen:16, DKLen:32}.
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:32768, R:8, P:4, SaltLen:16, DKLen:32} took 657.942582ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:16384, R:8, P:8, SaltLen:16, DKLen:32} took 495.989112ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:8192, R:8, P:17, SaltLen:16, DKLen:32} took 463.957228ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:4096, R:8, P:34, SaltLen:16, DKLen:32} took 520.679261ms (<nil>)
	scrypt_test.go:140: GenerateFromPassword with scrypt.Params{N:512, R:8, P:280, SaltLen:16, DKLen:32} took 527.884269ms (<nil>)
FAIL
exit status 1
FAIL	github.com/elithrar/simple-scrypt	20.183s
The command "go test -v ./..." exited with 1.

[LouisChrist]

It works perfectly fine on windows, but on linux not at all. The measured runtime of GenerateFromPassword() with the same parameters varies on my linux system by about 50ms to 100ms.

This may be related to #8.

[elithrar]

The problem is that in a CI, or other contended environment, a +/- 250ms difference on a 500ms target isn't that absurd given the "expense" of scrypt.

Calibration, being what it is, isn't perfect. Proposing:

• The 500ms tests should be more forgiving (+/- 75%)
• Add additional tests (e.g. 1500ms) to help with accuracy at higher absolute run times.

[LouisChrist]

This should work better. It's slower but the results are more consistent. I ran the tests 20 times and had no failures.

This might also fix #8. TestCalibrate fails around 80% percent of the time with the old Calibrate function on my computer.

[elithrar]

Almost done!

[elithrar]

Looks good, but one last ask: can you change the error message to detail what "around" means - e.g. "expected between {lower bound} and {upper bound}".

[LouisChrist]

Yes of course