john-users - On the usefulness of HT with compact SSE code

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

Date: Tue, 20 Jul 2010 19:12:35 +0200
From: bartavelle@...quise.net
To: john-users@...ts.openwall.com
Subject: On the usefulness of HT with compact SSE code

I just ran a distributed cracking job with my crypt-MD5 implementation. 
I used X5650 chips with and without HT. The "with HT" cracking only ran 
for about 3 hours, while the other ran for 10 days, but I believe this 
is representative.

I used markov mode (level 280) on a real password. My crypt-MD5 
implementation was compiled with icc 11 and is quite fast. Here are the 
per core speeds :

No HT : 30642.044 pwd/s
HT    : 16170.397 pwd/s
-test : 34700.000 pwd/s

HT gave a 5.5% speed increase. I believe this is due to the fact that 
the SSE code is quite effective. It might be possible to gain a bit by 
running 6 SSE tasks, and 6 "standard" tasks, but I didn't test this.