[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20110525203856.GA22015@openwall.com>
Date: Thu, 26 May 2011 00:38:56 +0400
From: Solar Designer <solar@...nwall.com>
To: john-dev@...ts.openwall.com
Subject: Re: BSDI poor in OMP4/OMP7 when doing real work
On Wed, May 25, 2011 at 10:18:38PM +0200, magnum wrote:
> What I don't understand though is why is the reported c/s is so high?
That's because John actually tries lots of candidate passwords - a lot
more than it does when it is not parallelized, because in that case the
correct passwords are detected sooner.
Sequential algorithm:
for each candidate {
for each salt {
hash the candidate
compare the hash against loaded hashes
remove any matching hashes
remove the salt if it no longer has hashes
}
}
Parallel algorithm:
for each bunch of candidates {
for each salt {
hash the candidates
compare the hashes against loaded hashes
remove any matching hashes
remove the salt if it no longer has hashes
}
}
In your case, every bunch of candidates probably has just one correct
password for each salt. Yet the entire bunch is hashed. So the larger
the bunch is, the worse your actual speed is. But not the c/s rate,
because this many combinations per second are actually tested.
If you don't have that many correct passwords in your candidates stream,
then the c/s rate will reflect the actual performance much better.
BTW, this is not limited to OpenMP; for many hash types, we're hashing
more than one candidate password at a time even with single-threaded
builds - to increase instruction level parallelism, to make use of SIMD,
to make more efficient use of caches, to save on function call overhead.
However, it's with OpenMP when the number of candidates to hash at once
may become huge. Oh, and even more so with the GPU patches.
Alexander
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.
