phc-discussions - Re: Non-temporal writes and uninitialized memory

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]

Message-ID: <CAOLP8p5Y2cB0QXKOEvwN04Mutmo83_9BFYXr=SkB7SoB_b=+og@mail.gmail.com>
Date: Fri, 18 Apr 2014 16:19:38 -0400
From: Bill Cox <waywardgeek@...il.com>
To: discussions@...sword-hashing.net
Subject: Re: Non-temporal writes and uninitialized memory

I've attached my code that demonstrates the difference in temporal writes
to initialized vs uninitialized memory, along with a Makefile that works in
64-bit Arch Linux.

Here's runtime data for "foo" and "bar".  Foo does no temporal writes to
DRAM, and Bar does only temporal writes to DRAM, but they do the same
computation.  This is for one iteration, which means they are writing to
uninitialized memory.  They each are reading the previous "block" of
memory, a pseudo-random prior block, doing a couple mixing instructions,
and writing to the output block.  They are processing 2 GiB of memory:

temporal_hack> time foo
counter 4016075129

real    0m0.361s
user    0m0.250s
sys     0m0.110s

temporal_hack> time bar
counter 4016075129

real    0m0.403s
user    0m0.300s
sys     0m0.100s

Foo, the version without non-temporal writes, is much faster!

For this next run, I edited foo.c and bar.c so that the outer loop iterates
10 times, rather than once.  The first iteration writes to uninitialized
memory, but the next 9 will write to initialize memory, and these
iterations dominate runtime:

temporal_hack> time ./foo
counter 3129101245

real    0m3.775s
user    0m3.673s
sys     0m0.100s

temporal_hack> time ./bar
time ./foobar
counter 3129101245

real    0m2.724s
user    0m2.620s
sys     0m0.103s

This time, bar, the version with temporal writes is much faster!

I also wrote a longer version that does not use temporal writes for the
first loop, but does use the for the next 9 iterations.  It's the file
called foobar.c, and foobar is the fastest of them all:

temporal_hack> time ./foobar
counter 3129101245

real    0m2.683s
user    0m2.577s
sys     0m0.107s

I think this can be used to make Yescript and Lyra2 faster.  TwoCats,
because it always fills new memory, will still hash more memory in a given
length of time, but Lyra2 and Yescript should be able to achieve higher
bandwidth using temporal writes in their second loops.

Bill

Content of type "text/html" skipped

Download attachment "temporal_hack.tgz" of type "application/x-gzip" (880 bytes)