john-users - RE: DES with OpenMP

Openwall GNU/*/Linux 3.0 - a small security-enhanced Linux distro for servers
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 21 Jan 2012 13:58:12 +0000
From: Alex Sicamiotis <alekshs@...mail.com>
To: <john-users@...ts.openwall.com>
Subject: RE: DES with OpenMP




> Date: Sat, 21 Jan 2012 15:15:50 +0400
> From: solar@...nwall.com
> To: john-users@...ts.openwall.com
> Subject: Re: [john-users] DES with OpenMP
> 
> On Sat, Jan 21, 2012 at 12:22:18AM +0000, Alex Sicamiotis wrote:
> > ... assuming (erroneously) that the code is identical... Apparently it has significant differences that speed up the OMP version and the nonOMP version is somewhat suboptimal (?).
> 
> In your case, yes.  In most other cases, it's the other way around.
> 
> Things should be more reasonable with GCC 4.3.
> 
> Alexander


Results are in (@...z / no x.org / 1.7.9 plain - no jumbo):

1. No-omp / ASM=0

GCC 4.3.4 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     4444K c/s real, 4444K c/s virtual
Only one salt:  4262K c/s real, 4262K c/s virtual

GCC 4.6.2 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     3370K c/s real, 3370K c/s virtual
Only one salt:  3265K c/s real, 3265K c/s virtual

ICC 12.1 (-fast -march=core2)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     4487K c/s real, 4487K c/s virtual
Only one salt:  4316K c/s real, 4316K c/s virtual

...so GCC 4.3.4 does great here but ICC slightly outperforms it.


2. OMP with NUM_THREADS=1

GCC 4.3.4 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     4131K c/s real, 4131K c/s virtual
Only one salt:  3780K c/s real, 3782K c/s virtual

GCC 4.6.2 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     4168K c/s real, 4168K c/s virtual
Only one salt:  3813K c/s real, 3813K c/s virtual

ICC 12.1 (-fast -march=core2)
Many salts:     4625K c/s real, 4625K c/s virtual
Only one salt:  4187K c/s real, 4187K c/s virtual

Almost identical performance for 4.3.4 / 4.6.2...


3. OMP with NUM_THREADS=2 

GCC 4.3.4 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     7932K c/s real, 3969K c/s virtual
Only one salt:  6268K c/s real, 3134K c/s virtual

GCC 4.6.2 (-O2 -march=nocona)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     7967K c/s real, 3983K c/s virtual
Only one salt:  6316K c/s real, 3158K c/s virtual

ICC 12.1 (-fast -march=core2)

Benchmarking: Traditional DES [128/128 BS SSE2-16]... DONE
Many salts:     8900K c/s real, 4450K c/s virtual 
Only one salt:  7153K c/s real, 3575K c/s virtual

Again, almost identical performance for 4.3.4 / 4.6.2... the biggest difference of 4.3.4 vs 4.6.2 in my case seems to be the non-OMP build which reaches almost ICC levels.