phc-discussions - Re: [PHC] Initial hashing function. Feedback welcome

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: Tue, 31 Dec 2013 09:23:21 +0400
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Initial hashing function. Feedback welcome

On Tue, Dec 31, 2013 at 08:23:57AM +0400, Solar Designer wrote:
> If your "pages" do not correspond to a whole number of physical pages,
> you'll run into false sharing between your threads when they happen to
> write into the same page.  This can truly kill performance, and this
> effect and its worst case extent won't be easily predictable.  Slowdowns
> by many orders of magnitude are possible.
> 
> http://en.wikipedia.org/wiki/False_sharing

Correction: normally, false sharing refers to accesses to the same cache
line rather than merely the same page.  However, I've seen a very
similar effect with same-page accesses by multiple threads, although
that was with 4 KiB pages (on x86).  I think this may have to do with
TLB implementations on specific CPUs.  I do not have a full
understanding of this, haven't seen it documented by others (for memory
accesses that are not to same cache line), don't know under what exact
conditions it will manifest (except that same-page accesses at least one
of which is a write, is a prerequisite), and whether it can be seen with
accesses to addresses that are far apart yet within the same huge page.

It is possible that my guess when I had run into this issue (outside of
PHC submissions context) was wrong.  Maybe the similar symptoms had a
different explanation.  But the workaround of moving the threads'
accesses to be to different pages worked well for me.

Alexander