Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 2 Feb 2012 06:41:46 +0400
From: Solar Designer <>
Subject: Re: Adding a new format

On Tue, Jan 31, 2012 at 11:41:27PM +0100, magnum wrote:
> On 01/31/2012 11:10 PM, Claudio Broglia wrote:
> > - would you suggest to crypt many keys at every pass, to gain speed, or
> > to stick with MAX_KEYS_PER_CRYPT set to 1?
> If you calculate one hash at a time (like in an OpenSSL version) there
> is no gain in rasing it.

Actually, there may be some minor gain from reduced function call
overhead (one call per N hashes vs. one call per 1 hash) and from
potentially improved locality of reference - especially if L1
instruction cache is direct-mapped or has low associativity - like what
we had on Alpha and like what we apparently have on AMD Bulldozer now.

Speaking of the latter, according to what I read Bulldozer's instruction
cache is 2-way, but shared for two cores in a module and with 1 way per
core - which I interpret as it being effectively direct-mapped.  If so,
we may get large slowdowns on certain builds/runs where pieces of code
from different source files, but used inside the cracking loop happen to
get overlapping cache tags.


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.