Date: Sun, 17 Nov 2013 03:46:46 +0400
From: Solar Designer <solar@...nwall.com>
To: Daniel Franke <dfoxfranke@...il.com>
Cc: discussions@...sword-hashing.net
Subject: Re: [PHC] Some updates on EARWORM

On Sat, Nov 16, 2013 at 02:39:36PM -0500, Daniel Franke wrote:
> Solar Designer <solar@...nwall.com> writes:
> > On Fri, Aug 23, 2013 at 09:44:18PM -0400, Daniel Franke wrote:
> >> * I've cut CHUNK_WIDTH from 4 to 2, leaving CHUNK_LENGTH at 128. Testing
> >>   on my workstation seems to indicate that neither the reduction in
> >>   internal parallelism nor the increased frequency of random memory
> >>   accesses results in any performance penalty.
> 
> I've actually changed my mind again since I wrote this: I'm going with
> CHUNK_WIDTH=4, CHUNK_LENGTH=64. The wider internal state makes certain
> proofs simpler.

OK.  I look forward to reading your proofs.

> > Is there a performance penalty with lower CHUNK_WIDTH or/and
> > CHUNK_LENGTH?  If so, how bad is it e.g. for 2 KB, 1 KB, 512 bytes?
> 
> There is. I don't have the figures handy. I'll retest this later today
> once I've gotten my code pushed.

How do you explain this performance penalty, given that you prefetch the
data?  Is it TLB misses?  Is it memory latency?  (The time to process
2 KB not enough to cover the full latency of the concurrent prefetch?
No, on a CPU this should be more than enough.)  Is it loop overhead?

What page size are you (or the OS) using for the arena[] allocation?
With 2 MB pages, 256 MB may fully fit in the TLB.

> My GPU code had a couple stupid bugs (no surprise at this point) that
> make these numbers completely bogus. The 7850 actually takes about 3x
> the time that the CPU does. It suffers from the same sort of bottlenecks
> that bcrypt does.

bcrypt is actually very different: it makes 32-bit accesses, whereas the
GPU is only able to fetch full cache lines from global memory (and with
huge latency).  So a bcrypt/GPU implementation trying to use global
memory ends up transferring lots of data that the code doesn't actually
need.  This is a reason why implementations of bcrypt on GPU that only
use local memory (and deliberately don't even try to use all computing
resources, because there's not enough local memory to use them all) end
up being faster.

scrypt as used in Litecoin, on the contrary, needs 128-byte chunks of
data, so is reasonably friendly to GPU implementations.  Along with a 2x
TMTO, it ends up being 10x+ faster on high-end GPUs than on CPUs.

EARWORM's memory access pattern is similar to scrypt's, although the
chunks are larger than Litecoin's and prefetching is possible.  EARWORM
is memory bandwidth bound, and GPUs generally have more memory bandwidth
than CPUs.  So with a proper implementation and sufficient parallelism
you should get EARWORM to run a few times faster on GPU than on CPU.

There's a reason why it might not work that way, though: the larger
chunks might fill up the GPU's local memory and caches before the GPU
chip's external bus is made full use of.  (Yes, this does become similar
to bcrypt using local memory and leaving most other resources idle as a
result.)  The workaround would be to fetch portions of chunks, process,
discard, fetch further portions, ... refetch, ...

> > Is this a defensive or offensive kind of implementation (if it were
> > finished, optimized, cleaned up, etc.)?  It sounds like you're computing
> > just one instance of EARWORM, but with some parallelism in it (albeit by
> > far not enough parallelism to use a GPU optimally), so I assume
> > defensive?  Anyhow, this doesn't tell us much about GPU attack speeds on
> > EARWORM.
> 
> Actually, this seems to be enough parallelism to use my (not very
> high-end) GPU optimially; though at this phase of experimentation you
> should take that claim with a grain of salt. Remember that each workunit
> already has considerable internal parallelism.

Yes, I'll take it with a grain of salt.  I think you're underestimating
just how much parallelism your GPU needs to cover the memory latency,
although your prefetch-friendly algorithm potentially helps GPUs a lot,
perhaps much more than it helps CPUs.  Are you somehow triggering
prefetches in your GPU code too?

According to
http://en.wikipedia.org/wiki/Comparison_of_AMD_graphics_processing_units#Southern_Islands_.28HD_7xxx.29_Series
a 7850 has 153.6 GB/s memory bandwidth.  This is 3x faster than
quad-channel DDR3-1600.

> It's a design goal of EARWORM that if the defender has GPUs available,
> he can use them just as effectively as the attacker can (even if they're
> not very effective for either side).

Sure.  I think it should be a common design goal for all PHC submissions
that whatever kind of hardware is not much more beneficial to attackers
than it may be to defenders.

When you make a password hashing scheme that targets solely memory
bandwidth, and it is used with memory sizes (soon to be) available on GPU
boards (or an efficient TMTO is possible), you should expect that GPUs
will run it a few times faster than CPUs - at least in optimized
attacks, where extra parallelism is available (many candidate passwords).

What _might_ save EARWORM from GPU attacks is that it actually targets
not only memory bandwidth, but also this 4 KB chunk size per instance -
which might not leave enough parallelism for a GPU (even in an attack
scenario) or might force the GPU to do the discard/refetch thing I
mentioned above (thereby wasting some of the GPU's memory bandwidth).

> Yes. Thanks for giving me the necessary kick in the rear to get working
> on this again :-)

You're very welcome.  I am happy that you're working on this. :-)

Alexander

Powered by blists - more mailing lists