now on Ryzen performance
freebsd.ed.lists at sumeritec.com
Thu Dec 27 02:53:19 UTC 2018
On Fri, 21 Dec 2018 17:40:49 +0100
Andrea Venturoli <ml at netfence.it> wrote:
> On 12/21/18 10:27 AM, Andrea Venturoli wrote:
> > On 12/4/18 12:21 PM, Erich Dollansky wrote:
> >> I just followed this mailing list and read some random articles.
> >> There seem to be no real stability problems which cannot be
> >> explained by other things like faulty motherboards etc.
> > Hello.
> > Just in case it helps others, I think I can conclude my new Ryzen 7
> > system works fine: uptime is almost a week, with high and low load
> > cycles (some idle time, some Poudriere, some server work, etc...).
> > I'm still running 11.2, BTW.
> Hello again.
> Now that I verified this system stability, I moved into measuring its
> performance and I think something is wrong here.
> A buildworld took many hours more than on any other system I'm
> managing (some even quite old), so I concluded some more
> investigation was needed. So I compared it to a (software-wise
> identical) system with a (7 year old) AMD Phenom II X4 965 (just
> because this was the CPU I had more easily available).
> According to UserBenchmark, a Ryzen 2700 should be 74% faster at
> single threading: if I interpret this correctly, it means total
> execution time should be 43% less...
> However, "time some_intensive_single_thread_task" showed the Ryzen
> completing the same work in 24% more time!!!
> Going to four threads (which is the maximum available on the Phenom)
> shows the same results.
> Of course using all 8 cores on the Ryzen takes 26% less time than
> using 4 on the Phenom, but that's not the point (and it's still slow).
> I have disabled SMT in the BIOS, so I didn't try with 16 threads.
> Having powerd running or not does not make any difference.
> When running "powerd -v" I see the Ryzen always working at 3.2GHz,
> which it's its base speed. It doesn't seem to use its "Max Boost
> Clock" of 4.1GHz. Is this to be expected? Is this feature only
> available in 12.0?
> Then again, if it just scaled linearly between the two frequencies it
> would still be much slower than I expected...
> Any hint?
> Is moving to 12 going to change this?
I am on FreeBSD 12.0 with a Ryzen 2700X and experience the same
problem. It is a different picture for several ports where really all
cores are used and things get real fast. It seems to be a problem of
the make system as specifying -j 16 leads to errors typically resulting
from build being started while the result required for this start is
not provided by another thread.
Aren't people with real servers using FreeBSD anymore?
More information about the freebsd-questions