lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 12 Jan 2014 13:14:53 +0400
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] escrypt memory access speed (Re: [PHC] Reworked KDF available on github for feedback: NOELKDF)

Bill, all -

On Sat, Jan 11, 2014 at 11:23:17PM +0400, Solar Designer wrote:
> On Sat, Jan 11, 2014 at 10:50:53PM +0400, Solar Designer wrote:
> > Back to 2 rounds of Salsa20:
[...]
> real    0m1.305s
> user    0m40.109s
> sys     0m1.242s
> 
> 49.37 GB/s

That was on 2x E5-2670 with AVX.  As an experiment, I've added OpenMP
support to -nosse and built it with "icc -mmic".  First, I tested that
it produces the same results on Xeon Phi 5110P at 32 threads - it does.
Then I increased p to 240, so that I can run 240 threads as this device
needs for optimal performance.  I also tuned r.  Turns out that r=8 is
optimal for this device (r=32 is much slower).

Xeon Phi 5110P, using 2 GB, r=8 p=240, 240 threads, Salsa20/2, abusing
scalar units for computation (no SIMD code implemented yet), no
prefetches, 10 hash computations:

real    0m 5.00s
user    17m 54.14s
sys     0m 18.27s

2*3*10*2^30/10^9/5.00 = 12.88 GB/s

Ditto with Salsa20/8:

real    0m 9.86s
user    37m 23.30s
sys     0m 20.83s

2*3*10*2^30/10^9/9.86 = 6.53 GB/s

These are some poor speeds. :-(

I guess it'd be better with SIMD (not trivial: need to bring 4x+ more
parallelism down to instruction level to use 512-bit SIMD, maybe with
p=960 or higher) and with prefetches, but I'm not sure by how much.

Xeon Phi 5110P has theoretical peak memory bandwidth of 320 GB/s, so
we're _very_ far from reaching it with this mostly unoptimized code.

If anyone wants to play with this more, let me know - I'd be happy to
provide remote access to this machine.

Alexander

Powered by blists - more mailing lists