On Saturday, November 26, 2011 11:39:23 PM Robert Haas wrote:
> On Sat, Nov 26, 2011 at 5:28 PM, Andres Freund <andres@anarazel.de> wrote:
> > On Saturday, November 26, 2011 09:52:17 PM Tom Lane wrote:
> >> I'd just as soon keep the fields in a logical order.
> >
> > Btw, I don't think the new order is necessarily worse than the old one.
>
> You forget to attach the benchmark results.
>
> My impression is that cache lines on modern hardware are 64 or 128
> *bytes*, in which case this wouldn't matter a bit.
>
> But testing is even better than guessing.
Being prodded like that I ran a very quick benchmark on my workstation.
Unfortunately that means I cannot work during the time which is why I kept it
rather short...
That machine has 2 E5520@2.27GHz which means 2(sockets) * 4(cores) *
2(threads) and 24GB of ram.
Data was initialized with: pgbench -h /tmp/ --unlogged-tables -i -s 20 pgbench
pgbench -h /tmp/ pgbench -S -j 16 -c 16 -T 60
origsnap: 92825.743958 93145.110901 93389.915474 93175.482351
reordersnap: 93560.183272 93613.333494 93495.263012 93523.368489
pgbench -h /tmp/ pgbench -S -j 32 -c 32 -T 60
origsnap: 81846.743329 81545.175672 81702.755226 81576.576435
81228.154119 81546.047708 81421.436262
reordersnap: 81823.479196 81787.784508 81820.242145 81790.263415
81762.421592 81496.333144 81732.088876
At that point I noticed I had accidentally run with a nearly stock config...
An even shorter run with a more approrioate config yielded:
pgbench -h /tmp/ pgbench -S -j 32 -c 32 -T 20
origsnap: 102234.664020 102003.449741 102119.509053 101722.410387
101973.651318 102056.440561
reordersnap: 103444.877879 103385.888808 103302.318923 103372.659486
103330.157612 103313.833821
Looks like a win to me. Even on this comparatively small machine.
Andres