Message-ID: <CAOLP8p5-uNUjLV8sAP2_dYjt-h-Ayw+UVcX0avnVDQM-9-AZng@mail.gmail.com>
Date: Sat, 5 Sep 2015 05:07:39 -0700
From: Bill Cox <waywardgeek@...il.com>
To: "discussions@...sword-hashing.net" <discussions@...sword-hashing.net>
Subject: Re: [PHC] Low Argon2 performance in L3 cache

On Fri, Sep 4, 2015 at 5:11 PM, Solar Designer <solar@...nwall.com> wrote:

> Hi Bill,
>
> On Fri, Sep 04, 2015 at 04:51:31PM -0700, Bill Cox wrote:
> > Argon2d memory for 1.2ms hash: 2200 KiB
> > Serial multiplies: 2200*96 = 211,200
> > ASIC attacker speed using 1ns multipliers: 0.211ms
> > area-time product: 0.465 s-KiB
> >
> > TwoCats memory for 1.2ms hash: 8192 KiB
> > Serial multiplies: 526336
> > ASIC attacker speed using 1ns multipliers: 0.526ms
> > area-time product: 4.31 s-KiB
> >
> > It looks like TwoCats will have about 9X improved time-area defense, when
> > we take into account the multiplication chains.
>
> What is it that makes Argon2d so much slower?  Is it needing to perform
> two BLAKE2b rounds per sub-block, and the intermediate writes to state?
>

Mostly 2 things: Too many Blake2 rounds, and having state that does not fit
into the mmx registers.  Cutting the Blake2 rounds in half looks fairly
simple, but I don't know what to do about the state variables.  Argon2d
uses 1KiB of state variables, and this is equal to it's block size.  If I
decreased the state size, RAM latency issues would begin to dominate.


> Is memory (de)allocation overhead excluded from the 1.2ms for both of
> these?  And no zeroization done either?  At least we need to ensure the
> benchmarks are consistent in this respect.
>

No, and I think there's a C++ extra zeroing going on in Argon2d.  All three
are allocating and deallocating memory.


> Can you tune Argon2d and TwoCats for same defensive throughput per CPU
> chip (with multiple independent concurrent instances), rather than for
> same defensive latency, for a comparison like this?
>

This would be ideal, in that having only password hashing running on a CPU
solves problems related to how these algorithms do not work and play well
with other services on the same CPU.  This also improves defense against
side-channel attacks.  I think you well optimized Yescrypt for this use
case.  However, there are complications in treating one service different
from another.  Data center guys don't like it.  It would improve defense,
though.

My machine has 12 MiB of cache.  This would allow 3 instances to run
without harming each other.  I verified this at one point.  The interaction
is in the noise, maybe slowing each other down by 1-3%.  I have 6 CPU cores
,so a 2MiB hash size would let me do 6 in parallel.  Eliminating the memory
initialization would help further.  I would need to see if at that speed
they still do not fight over a common resource like total L3 bandwidth.  I
would not like dropping to 2MiB, though, as defense does seem to go as the
square of memory.

So, 3 4MiB threads run at almost the same rate as 1.  Just multiply by 3...
this is about 4,000 4MiB hashes per second for TwoCats, I think.  I'd have
to verify that...  If Argon2's speed issues look difficult to overcome for
a 4MiB hash, I should probably look into Yescrypt for this particular space.

I think it's primarily throughput per chip that matters at memory sizes
> and low latencies like this.  It doesn't really matter if it takes 1ms
> or 2ms of latency to reach a few MB, but it does matter what memory per
> hash you can reach within a given hashes per second budget (e.g. for
> 5000 per second per chip).
>
> Alexander
>

I agree.  Ideally, data centers would allocate entire CPUs to this task, so
hashes per second per CPU, hash size, and the various defenses are what
count.

Bill

Content of type "text/html" skipped

Powered by blists - more mailing lists