Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <d36b6c4fd0a0c5f1aed5444ddceb1934@smtp.hushmail.com>
Date: Wed, 02 Oct 2013 02:55:18 +0200
From: magnum <john.magnum@...hmail.com>
To: john-dev@...ts.openwall.com
Subject: Re: X32 build, anyone?

I built latest snapshot OpenSSL for X32 (no separate patch needed) and 
did some testing with bleeding-jumbo. Just like with core, there were no 
build problems and all formats pass self-tests as well as Test Suite.

Benchmarks are generally barely faster than x86_64 - so a lot faster 
than 32-bit. Real tests show the same.

On 2013-05-22 12:39, Solar Designer wrote:
> There may be slight performance gains when going from x86_64 to x32
> for:
>
> 1. Lots of fast hashes (thousands or millions) loaded for cracking,
> where the savings from smaller pointer size may reduce the working
> set size and thus result in more cache hits.  However, this effect
> should be reduced in bleeding (as compared to 1.7.9-jumbo-7) due to
> us starting to use bitmaps (instead of hash tables), which are same
> size and speed regardless of pointer size.

I tried this, using various subsets of the 132M MD5 hashes from 
InfoSecSouthwest2012_Ripe_Hashes.tgz. For some reason I saw very little 
gain, if any, over an x86_64 build.

While a 32-bit build can't load files larger than 2 GB, the X32 build 
was only limited by the 4 GB virtual memory size, so the latter could 
load more hashes at once but not all 132M of them. BTW for these tests I 
bumped PASSWORD_HASH_SIZE_FOR_LDR to max.

> 2. Bitslice DES in OpenMP-enabled builds, where DES_bs_all structs
> contain arrays of pointers, and together these structs don't fit in
> L1 data cache.  (They could fit with fewer instances computed per
> thread, but then the OpenMP overhead would be higher.)  Without
> OpenMP, we have just one DES_bs_all, which fits in L1 data cache, so
> there shouldn't be much difference.

I tested this briefly using 10 loaded hashes. The X32 build was not 
faster than x86_64, in fact it was slower within a percent (just like 
for non-OMP).

I'm done with this for now. I haven't seen any reason to go from x86_64 
to X32 but it does work just fine.

magnum

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.