[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 26 May 2015 20:22:50 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: interleaving in SHA256 & SHA512
On 2015-05-26 09:26, Lei Zhang wrote:
> On May 25, 2015, at 4:05 PM, magnum wrote:
>> BTW for interleaving factors, you should not just try multiples
>> like 2, 4, 8. You need to try all of 2, 3, 4, 5 - and more than 5
>> is probably never worthwhile (if it is, I guess it indicates we
>> should look into the code instead of just bumping para). So maybe
>> you need to retry all those benchmark yet again ;-)
>
> Ok, here's a re-test on MIC, with both a OpenMP-enabled and a OpenMP-disabled build:
> sha512crypt
> x1
> Raw: 6262 c/s real, 26.1 c/s virtual
> x2
> Raw: 6606 c/s real, 27.9 c/s virtual
> x3
> Raw: 6658 c/s real, 28.1 c/s virtual
> x4
> Raw: 7029 c/s real, 29.6 c/s virtual
> x5
> Raw: 6946 c/s real, 30.3 c/s virtual
Not a lot, but indeed a little RoI. I was starting to think all this
work was for no good :-)
Like I said on GitHub, sha512crypt is a bit tricky because it sorts
lengths in order to avoid diverging threads. So there's not only
SIMD_PARA and OMP_SCALE but also SIMD_COEF_SCALE and all three will
affect each other to some degree.
Maybe when all SHA2 interleaving is stable we can find some better
format for benchmarking, eg. pbkdf2-hmac-sha512. I have a feeling
finding a perfect SIMD_PARA might involve decreasing a previously tuned
OMP_SCALE.
magnum
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
