Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 25 Jul 2015 18:57:57 +0200
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: PHC: my yescrypt and lyra2 benchmarks

On Sat, Jul 25, 2015 at 06:40:13PM +0200, Agnieszka Bielec wrote:
> 2015-07-25 18:37 GMT+02:00 Agnieszka Bielec <bielecagnieszka8@...il.com>:
> > so far I'm getting (for r=6) (after I removed copying and added
> > vectors to pwxform)
> > [a@...er run]$ ./john --test --format=yescrypt-opencl --dev=5
> > Benchmarking: yescrypt-opencl [Salsa20/8 OpenCL (inefficient,
> > development use only)]... Device 5: GeForce GTX TITAN
> > memory per hash : 1.51 MB
> > DONE
> > Speed for cost 1 (N) of 2048, cost 2 (r) of 6, cost 3 (p) of 1, cost 4
> > (t) of 0, cost 5 (g) of 0
> > Raw:    1003 c/s real, 1008 c/s virtual
> 
> for r=8:
> 
> [a@...er run]$ ./john --test --format=yescrypt-opencl --dev=5
> Benchmarking: yescrypt-opencl [Salsa20/8 OpenCL (inefficient,
> development use only)]... Device 5: GeForce GTX TITAN
> memory per hash : 2.10 MB
> DONE
> Speed for cost 1 (N) of 2048, cost 2 (r) of 8, cost 3 (p) of 11, cost
> 4 (t) of 0, cost 5 (g) of 0
> Raw:    718 c/s real, 718 c/s virtual

These speeds look good to me: they're consistent with what others
reported for specialized code for BSTY mining.  Well, the latter is
moderately worse - by 30% or so - but it's reasonable given that you
don't yet use inline ptx asm, and they do.

Please prepare a separate posting with Lyra2 and yescrypt benchmarks
with the parameters I requested a few days ago, suitable for sharing it
with the PHC community.

BTW, the "inefficient, development use only" comments we use in two
cases: for "fast hashes", where we don't yet have on-GPU candidate
password generation (so speeds are like 100x worse than those of
competing crackers) (*), and for formats where we know we have a faster
alternative (e.g., sha512crypt-cuda has this comment, advising people to
use sha512crypt-opencl instead).  I think you shouldn't put it on your
PHC finalist formats.  Perhaps you inherited it from building your GPU
formats based on a "fast hash" format (which magnum and I previously
told you was a bad idea).

(*) Sayantan's recent commits correct this for "mask mode" for a handful
of "fast hash" types.

Alexander

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.