phc-discussions - Re: [PHC] Re: Updated tests document (version 2)

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: <553F1E7E.9010807@dei.uc.pt>
Date: Tue, 28 Apr 2015 06:45:34 +0100
From: Samuel Neves <sneves@....uc.pt>
To: discussions@...sword-hashing.net
Subject: Re: [PHC] Re: Updated tests document (version 2)

On 04/28/2015 12:55 AM, Bill Cox wrote:
> I think there are just 3 candidates now supporting multiple threads.  It is
> surprisingly hard to guestimate parallel thread performance from parallel
> process performance.  The reasons remain mysterious to me, but the short
> version is that each thread seems to want sole access to it's own page.
> Mixing nearby read/writes between threads thrashes something, maybe the
> translation lookaside buffer?

The more common (and costly) stall when reading and writing nearby memory accesses is false sharing: Core 0 writes to
cache line X, while Core 1 reads from the same cache line X. Since our chips are cache-coherent, this forces X to be
constantly flushed to main memory, even if Core 0 and 1 never actually accessed the same memory address.

Cores may also be fighting for control of the outermost cache layer (L3 these days). Each core has its own TLB, and
presumably every page is setup ahead of time, so I'm not sure that's the main factor in such slowdowns. But as always,
measuring these performance events is the only way to be sure.