On Fri, Apr 12, 2019 at 4:23 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> ... if we also prefetch during
> the build phase ...
Here's an experimental patch to investigate just that part. I tried
initiating a prefetch of the bucket just before we copy the tuple and
then finally insert it, but it doesn't seem to be far enough apart (at
least for small tuples), which is a shame because that'd be a one line
patch. So I made a little insert queue that prefetches and defers the
insertion until N tuples later, and then I managed to get between 10%
and 20% speed-up for contrived tests like this:
create unlogged table t as select generate_series(1, 100000000)::int i;
select pg_prewarm('t');
set work_mem='8GB';
select count(*) from t t1 join t t2 using (i);
master patched/N=1 patched/N=4
workers=0 89.808s 80.556s (+11%) 76.864 (+16%)
workers=2 27.010s 22.679s (+19%) 23.503 (+15%)
workers=4 16.728s 14.896s (+12%) 14.167 (+18%)
Just an early experiment, but I though it looked promising enough to share.
