[<prev] [next>] [day] [month] [year] [list]
Message-ID: <a1739247-4c90-d9f1-b2b5-4989363acc7e@openwall.net>
Date: Fri, 3 Feb 2017 10:29:06 -0600
From: jfoug <jfoug@...nwall.net>
To: john-users@...ts.openwall.com
Subject: Re: to Single or not to Single
On 2/3/2017 6:14 AM, Solar Designer wrote:
> One obvious reason for a lower speed is that you're presumably testing
> only 1 password per salt, whereas we're optimizing for many passwords to
> be tested at once. On a non-OpenMP build with AVX:
>
> $ ./john --list=format-all-details --format=dynamic_25 | fgrep keys
> Min. keys per crypt 64
> Max. keys per crypt 1680
For AVX2 you will get 8X*PARA x performance loss (plus much other
overhead loss due to dyna being very heavy on extraneous CPU usage).
The 1680 is actually used to absorb and hide some of that additional
overhead. The 64 min was due to getting a value that worked for all
hash types. It really 'could' be larger, but making that adjustment is
very tricky, since we have to deal with all possible SIMD coefs, AND
SIMD para for all different hash types which dyna handles. But you are
VERY right, single does have some very huge hits in this type usage
(which can be an ideal usage), due to the fact that the blocks being
tested are not being filled up optimally. Even though min is 64, the
code will not be forced into doing 64 hashes. It will stop when the
number of ready to go passwords (rounded up to the next simd*para) is
enough. So if there is only 1 password, then only 1 call to the
simd_crypt() will be done (which on AVX2 is 8x*para)
NOTE, this type run should NEVER be done with OMP, or with fork. It
will simply increase the loss. One thing that 'may' improve the
throughput is to actually not do the SingleRetestGuessed=Y flag Because
in that way, the 1 'excellent' candidate will get buffers filled in with
other cracked data. If run like this you will certainly want to run
some smaller 'testing' cases to make sure it does not cause the overall
run to increase in time much, and actually see if there is a decent
improvement in final end percentages cracked. For the AVX2, if you add
7 more passwords per each crack call (15 more is build is para=2), the
program will run through the data at exactly the same rate, BUT all
those other 'possible' passwords will be tested for nearly free.
Jim.
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.
