pgbench performance is lagging compared to Linux and DragonflyBSD?

[Garrett Cooper]

On Mon, Nov 5, 2012 at 10:52 AM, Yuri <yuri at rawbw.com> wrote:

> There is the post by DragonflyBSD folks that claims that Linux and
> DragonflyBSD are quite ahead of FreeBSD on pgbench test on 12 Core 2x Xeon
> X5650 with 24 threads.
> Here are their results with graphs: http://lists.dragonflybsd.org/**
> pipermail/users/attachments/**20121010/7996ff88/attachment-**0002.pdf<http://lists.dragonflybsd.org/pipermail/users/attachments/20121010/7996ff88/attachment-0002.pdf>
> And here is their original post: http://lists.dragonflybsd.org/**
> pipermail/users/2012-October/**017536.html<http://lists.dragonflybsd.org/pipermail/users/2012-October/017536.html>
>
> I am not sure if this is the problem of some sysctl or kernel parameters
> or some serious system issue.
>
> It looks like the DragonflyBSD folks made a goal to do well on pgbench and
> got to the level of ~88% of linux with 80 clients.


The important item that has been left out (or is just implied as OS level
defaults) is sysctl/tunable variables set in the *BSD OSes (on DFly,
FreeBSD, and NetBSD). Unfortunately (based on my experience) FreeBSD could
be a lot better when it comes to defaults, and more tuning is required to
get better performance. So if they're working with the OS defaults, this
might not be a fair equivalent to the best performance that FreeBSD can
yield, but it's probably fair to do this for the sake of repeatability and
to prove what these OSes can do out of the box. This is in addition to the
[lock] contention issues that jeffr@ and a few others are working on
alleviating.

FWIW, I think that the last time scheduler benchmarks from anyone at
@FreeBSD.org (was kris@ the last one, or has flo@ run benchmarks since
then? My Googling is a bit inconclusive) was run was several years ago as
well, so if Linux has improved I'm not at all surprised. However, please
also take into consideration that the hardware then and the hardware now
are grossly different. So the interactions between the hardware then and
the hardware now might differ greatly. In short, more inspection needs to
be done to figure out whether or not the findings are true [with caveats]
or false.

Thanks,
-Garrett