Date: Thu, 17 Apr 2014 17:10:49 -0400
From: Bill Cox <waywardgeek@...il.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Lyra2 initial review

On Thu, Apr 17, 2014 at 4:40 PM, Marcos Antonio Simplicio Junior <
mjunior@...c.usp.br> wrote:

> Things are getting clearer. Thank you :-)
>
> *De: *"Solar Designer" <solar@...nwall.com>
> *Para: *discussions@...sword-hashing.net
> *Enviadas: *Quinta-feira, 17 de Abril de 2014 15:53:15
>
> *Assunto: *Re: [PHC] Lyra2 initial review
>
> On Thu, Apr 17, 2014 at 01:42:01PM -0300, Marcos Antonio Simplicio Junior
> wrote:
> > Actually an explanation would be really good, because I confess that,
> except for "memory bandwidth", I'm not sure I understand what those numbers
> mean or how to computed them...
>
> Does the "memory bandwidth" figure for Lyra2 look correct to you, given
> the "time" output Bill posted?
>
> Well, the "time" seems OK, although it is quite worse than what we got in
> our machine. The number "7" being multiplied by "Peak memory/second" eludes
> me, though. Each hashing operation in Lyra2 takes as input two memory
> locations ("prev" and "row*")  and then its output is written to two memory
> locations ("row" and "row*). During Setup, "row" is just written over,
> while "row*" is XORed with the output (which may count as a "read"?);
> during the Wandering phase, both "row" and "row*" are XORed with the output.
>
> Am I missing something? The r/w on the "state" variable, maybe?
>

I'm not sure where the disconnect is.  I measured Lyra2 hashes 2

>From the v1 version of the Lyra2 paper I just downloaded from the PHC site,
I see that the fastest hashing time reported is for T=1, the same as I ran.
 From the chart on the top of page 40, it looks like it hashes 1600 MiB in
about 1.5 seconds, or about 1.04 GiB/s.  I reported Lyra2 running on my
machine at 1.52 GiB/s.  It seems my machine is faster :-)

Converting from MiB in 1.5 seconds to GiB/s assumes that the runtime is
proportional to the amount of memory hashed, which is a reasonable
approximation at this size of memory.  However, it may be simpler to
compute the GiB/s of DRAM bandwidth directly from the measured data.

Lyra2 hashed 2 GiB of memory in 1.32 seconds on my machine.  On average,
every location in external DRAM was accessed (read or written) about 7
times.  Lyra2 does 9 total memory accesses, but 2 of them are to read the
previously written row, which I expect to still be in cache, so I don't
count them.  To get the average memory bandwidth for the whole run, I just
multiply the memory size by 7 and divide by the runtime:

bandwidth = 2GiB*7/1.32 = 10.6 GiB/s

This matches up well with the speed record, set by EARWORM, of a bit more
than 12 GiB/s.  At this bandwidth, Lyra2 is clearly memory speed limited on
my machine.

> nCols is set to 64 by default, which means that each row has 64*768 bits =
> 6144 bytes = 6 KiB. In our machine we got better results with nCols = 128
> (12 KiB), though. Those are the figures we put in the specification.
>
> I found that 16 KiB block size for TwoCats is quite a bit faster than
smaller sizes.  Dropping to 4KiB slows it down by 16%.  I think this is the
main remaining difference in memory bandwidth between TwoCats and Yescript.
 Alexander uses 1KiB block sizes.  That's better for GPU defense, and I'm
not surprised Alexander prefers this trade-off.  Being mostly unfamiliar
with GPUs, I cringe at weakening defense against ASIC attacks, and prefer
the 16KiB size for its speed.

Bill

Content of type "text/html" skipped

Powered by blists - more mailing lists