Openwall GNU/*/Linux - a small security-enhanced Linux distro for servers
[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Date: Mon, 8 Aug 2011 15:22:44 +0200
From: Lukas Odzioba <lukas.odzioba@...il.com>
To: john-dev@...ts.openwall.com
Subject: Lukas's Status Report - #13 of 15

Last week I've:
-optimized phpass,sha-crypts patches
-retested them before contest
-cracked some passwords with john-users team
-fixed 2 GPU problem on my system, I thougth that it was drivers fault
but problem laid in 32bit kernel. Adding vmalloc=256MB fixed X's
problem.


This week:
-fix/test OpenCL build problems
-test phpass more extensively
-work on previous (unfinished) patches

Thoughts:
I think that from now it will be better keep cuda patches together
like in "allcuda" patch, because it is easier to manage one version. I
couldn't find good name for it when I was uploading to the wiki, but
now I think it "slowcuda" is better name.
Today I've did some changes on phpasscode and got 5% more power from
it. I had to duplicate whole MD5 block but it is not using any
external memory during main processing.

I found on Ivan Golubev's blog
(http://www.golubev.com/blog/?tag=gtx460) interesting information
about GF104 and newer nvidia chips saying:

"But main thing that youíre need to vectorize your code for GF104 (and
106/108) to get full performance of GPU as it already happens with ATI
GPUs. Processing single hash per thread will drop performance by 2/3
(for algorithms which instructions heavily depends on each other, like
MD5), thus making 336SP of GTX460 looking like only 224SP. There are a
lot of another changes at GF104 but they arenít so important for hash
calculations like code vectorization."

That gives me hope for further cuda patches optimizations in the near future.


>From benchmarks posted on ml of GTX480 it's clear that phpass code is
alu-bounded, and scales great on more powerfull graphics cards, while
MD5-crypt and SHA-crypts still have got memory bottleneck which I must
get rid off, or at least minimize. This explain why modyfing MD5's F()
and G() functions did not help at md5crypt exactly as Solar stated
earlier (but helped for phpass).

Groszek and Solar reported problems with phpass used during the
contest, I need to investigate it too.

Lukas

Powered by blists - more mailing lists

Your e-mail address:

Powered by Openwall GNU/*/Linux - Powered by OpenVZ