this 48-core box...

[Adrian Chadd]

The cache alignment happens because it hits a specific size threshold, and
jemalloc/phkmalloc(I think!) just round everything up to be page size
aligned.

The underlying problem may actually be a code change to how the math is
done. It just runs slower on page-aligned alignments..

adrian



On 22 September 2013 05:10, Eduardo Morras <emorrasg at yahoo.es> wrote:

> On Sat, 21 Sep 2013 12:53:36 -0700
> Adrian Chadd <adrian at freebsd.org> wrote:
>
> > .. just as a data point - there was a thread a while ago about numeric
> > processing performance on linux vs bsd.
> >
> > It all boiled down to how jemalloc versus the linux allocator(s) allocate
> > blocks. jemalloc will page align things after a certain size. Linux
> didn't.
> > So when doing numeric processing, there was a lot of cache aliasing going
> > on leading to inefficient cache usage and redundant memory operations.
> >
> > When the same workload on Linux was run on FreeBSD but with the Linux
> > library/allocators, the performance was identical.
> >
> > No-one followed through. I think I may have to write a blog post about
> it.
>
> There's no MALLOC_OPTIONS flag to set/unset this, but adding a new flag to
> disable a feature is easier (or should be) than implementing new one. The
> only problem I see to this is if the cache align happens at sbrk/mmap level.
>
>
> >
> > -adrian
> > _______________________________________________
> > freebsd-questions at freebsd.org mailing list
> > http://lists.freebsd.org/mailman/listinfo/freebsd-questions
> > To unsubscribe, send any mail to "
> freebsd-questions-unsubscribe at freebsd.org"
>
>
> ---   ---
> Eduardo Morras <emorrasg at yahoo.es>
> _______________________________________________
> freebsd-questions at freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-questions
> To unsubscribe, send any mail to "
> freebsd-questions-unsubscribe at freebsd.org"
>