[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20130527180046.XEA1N.185069.imail@eastrmwml205>
Date: Mon, 27 May 2013 18:00:46 -0400
From: <jfoug@....net>
To: john-dev@...ts.openwall.com
Subject: Re: OMP for raw formats
---- magnum <john.magnum@...hmail.com> wrote:
> On 27 May, 2013, at 21:39 , magnum <john.magnum@...hmail.com> wrote:
> > Jim recently enabled OMP for the fast raw formats and that made me enable it for nt2 as well (it was there but disabled). They all work fine on Intel i7 but I see massive performance regressions on AMD hardware, not only XOP/Bull but also older AMD. Even using OMP_NUM_THREADS=1 the speed is terrible. I think we need to revert this or figure out some way to handle it.
>
> I see now Bull's results are better than the older AMD six-core I tried earlier. That one really had awful regression at OMPx1, Bull does not - it just doesn't take off using more cores, which is sad but not that much of a problem.
I just did my large inc test on raw-md5, and on my very old, slow 2x intel i3 laptop, the speed went from 16.5m to 17.5m, so there 'is' some gain to be had. I will check in the changes (pretty minor). Lets see how they behave on older AMD, if it makes any difference, and if the difference is enough to overcome the earlier objections you were having.
Jim.
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
