Re: PostgreSQL performance on FreeBSD

On Tue, Aug 12, 2014 at 1:09 PM, John Baldwin <jhb_at_freebsd.org> wrote:

> On Wednesday, July 16, 2014 1:52:45 pm Adrian Chadd wrote:
> > Hi!
> >
> >
> > On 16 July 2014 06:29, Konstantin Belousov <kostikbel_at_gmail.com> wrote:
> > > On Fri, Jun 27, 2014 at 03:56:13PM +0300, Konstantin Belousov wrote:
> > >> Hi,
> > >> I did some measurements and hacks to see about the performance and
> > >> scalability of PostgreSQL 9.3 on FreeBSD, sponsored by The FreeBSD
> > >> Foundation.
> > >>
> > >> The results are described in https://kib.kiev.ua/kib/pgsql_perf.pdf.
> > >> The uncommitted patches, referenced in the article, are available as
> > >> https://kib.kiev.ua/kib/pig1.patch.txt
> > >> https://kib.kiev.ua/kib/patch-2
> > >
> > > A followup to the original paper.
> > >
> > > Most importantly, I identified the cause for the drop on the graph
> > > after the 30 clients, which appeared to be the debugging version
> > > of malloc(3) in libc.
> > >
> > > Also there are some updates on the patches.
> > >
> > > New version of the paper is available at
> > > https://www.kib.kiev.ua/kib/pgsql_perf_v2.0.pdf
> > > The changes are marked as 'update for version 2.0'.
> >
> > Would you mind trying a default (non-PRODUCTION) build, but with junk
> > filling turned off?
> >
> > adrian_at_adrian-hackbox:~ % ls -l /etc/malloc.conf
> >
> > lrwxr-xr-x  1 root  wheel  10 Jun 24 04:37 /etc/malloc.conf -> junk:false
> >
> > That fixes almost all of the malloc debug performance issues that I
> > see without having to recompile.
> >
> > I'd like to know if you see any after that.
>
> OTOH, I have actually seen junk profiling _improve_ performance in certain
> cases as it forces promotion of allocated pages to superpages since all
> pages
> are dirtied.  (I have a local hack that adds a new malloc option to
> explicitly
> memset() new pages allocated via mmap() that gives the same benefit without
> the junking overheadon each malloc() / free(), but it does increase
> physical
> RAM usage.)
>
>

John,

A couple small steps have been taken toward eliminating the need for this
hack: the addition of the "page size index" field to struct vm_page and the
addition of a similarly named parameter to pmap_enter().  However, at the
moment, the only tangible effect is in the automatic prefaulting by
mmap(2).  Instead of establishing 96 4KB page mappings, the automatic
prefaulting establishes 96 page mappings whose size is determined by the
size of the physical pages that it finds in the vm object.  So, the
prefaulting overhead remains constant, but the coverage provided by the
automatic prefaulting will vary with the underlying page size.