On 2014-06-30 19:44:47 +0200, Andres Freund wrote:
> On 2014-06-30 13:15:23 -0400, Robert Haas wrote:
> > People who write code that uses atomics are not
> > likely to think about how those algorithms will actually perform when
> > those atomics are merely emulated, and I suspect that means that in
> > practice platforms that have only emulated atomics are going to
> > regress significantly vs. the status quo today.
>
> I think you're overstating the likely performance penalty for
> nonparallel platforms/workloads here quite a bit. The platforms without
> changes of gettings atomics implemented aren't ones where that's a
> likely thing. Yes, you'll see a regression if you run a readonly pgbench
> over a 4 node NUMA system - but it's not large. You won't see much of
> improved performance in comparison to 9.4, but I think that's primarily
> it.
To quantify this, on my 2 socket xeon E5520 workstation - which is too
small to heavily show the contention problems in pgbench -S - the
numbers are:
pgbench -M prepared -c 16 -j 16 -T 10 -S (best of 5, noticeably variability)
master: 152354.294117
lwlocks-atomics: 170528.784958
lwlocks-atomics-spin: 159416.202578
pgbench -M prepared -c 1 -j 1 -T 10 -S (best of 5, noticeably variability)
master: 16273.702191
lwlocks-atomics: 16768.712433
lwlocks-atomics-spin: 16744.913083
So, there really isn't a problem with the emulation. It's not actually
that surprising - the absolute number of atomic ops is prety much the
same. Where we earlier took a spinlock to increment shared we now still
take one.
I expect that repeating this on the 4 socket machine will show a large
gap between lwlocks-atomics and the other two. The other two will be
about the same, with lwlocks-atomics-spin remaining a bit better than
master.
Greetings,
Andres Freund
-- Andres Freund http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services
