john-users - Re: Anyone want to benchmark AVX2 code for bcrypt

Follow @Openwall on Twitter for new release announcements and other news

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20130713192024.GA25369@openwall.com>
Date: Sat, 13 Jul 2013 23:20:24 +0400
From: Solar Designer <solar@...nwall.com>
To: john-users@...ts.openwall.com,
	"Sc00bz64@...oo.com" <sc00bz64@...oo.com>
Subject: Re: Anyone want to benchmark AVX2 code for bcrypt

On Wed, Jun 26, 2013 at 09:09:27AM -0700, Sc00bz64@...oo.com wrote:
> Windows 8 x64 4770k at 4.1GHz
> 
> Single threaded performance:
> AVX2: 868.6 h/s
> Hashcat: 1170 h/s (hashcat-cli64.exe -m 3200 -a 3 -n 1 m3200.txt -1 ?l?u?d?s ?1?1?1?1?1?1?1)

4770K at stock clocks (I think it's 3.5 GHz base, 3.9 GHz max turbo),
Ubuntu 12.04.2, your original binary:

solar@...l:~/j/bcrypt/bcryptavx2$ ./bcryptbench64.orig
Tests PASSED
Benchmarking...
AVX2: 8*256 took 1.9838 sec (1032.4 h/s) 0
Normal: 1024 took 1.6849 sec (607.8 h/s) 0

My rebuild of it:

solar@...l:~/j/bcrypt/bcryptavx2$ ./bcryptbench64
Tests PASSED
Benchmarking...
AVX2: 8*128 took 1.0204 sec (1003.5 h/s) 0
Normal: 1024 took 1.6829 sec (608.5 h/s) 0
solar@...l:~/j/bcrypt/bcryptavx2$ gcc --version
gcc (GCC) 4.9.0 20130707 (experimental)

JtR using one core:

solar@...l:~/j/john-1.8.0/run$ ./john -te -form=bcrypt
Benchmarking: bcrypt ("$2a$05", 32 iterations) [Blowfish 32/64 X2]... DONE
Raw:    1110 c/s real, 1110 c/s virtual

... and in an OpenMP-enabled build:

solar@...l:~/j/john-1.8.0/run$ OMP_NUM_THREADS=1 ./john -te -form=bcrypt
Warning: OpenMP is disabled; a non-OpenMP build may be faster
Benchmarking: bcrypt ("$2a$05", 32 iterations) [Blowfish 32/64 X2]... DONE
Raw:    1104 c/s real, 1106 c/s virtual

JtR using all four cores:

solar@...l:~/j/john-1.8.0/run$ ./john -te -form=bcrypt
Will run 8 OpenMP threads
Benchmarking: bcrypt ("$2a$05", 32 iterations) [Blowfish 32/64 X2]... DONE
Raw:    6595 c/s real, 825 c/s virtual

> So not using AVX2 is faster.

Yes, but by very little.  I think we should be able to repair that by
tweaking the code.

I don't know why a bigger slowdown with AVX2 was seen on Windows.
Different compiler?

I also tried JtR bcrypt-opencl with both AMD's and Intel's OpenCL SDK on
this CPU.  AMD's produces 5120 c/s.  Intel's produces 2576 c/s (might be
using AVX2 and exceeding the cache).

Alexander

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.