On Tue, Jan 21, 2020 at 6:20 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> On Fri, Jan 10, 2020 at 1:52 PM Deng, Gang <gang.deng@intel.com> wrote:
> > Thank you for the comment. Yes, I agree the alternative of using '(!parallel)', so that no need to test the bit.
Willsomeone submit patch to for it accordingly?
>
> Here's a patch like that.
Pushed. Thanks again for the report!
I didn't try the TPC-DS query, but could see a small improvement from
this on various simple queries, especially with a fairly small hash
table and a large outer relation, when many cores are probing.
(Off topic for this thread, but after burning a few hours on a 72-way
box investigating various things including this, I was reminded of the
performance drop-off for joins with large hash tables that happens
somewhere around 8-16 workers. That's because we can't give 32KB
chunks out fast enough, and if you increase the chunk size it helps
only a bit. That really needs some work; maybe something like a
separation of reservation and allocation, so that multiple segments
can be created in parallel while respecting limits, or something like
that. The other thing I was reminded of: FreeBSD blows Linux out of
the water on big parallel hash joins on identical hardware; I didn't
dig further today but I suspect this may be down to lack of huge pages
(TLB misses), and perhaps also those pesky fallocate() calls. I'm
starting to wonder if we should have a new GUC shared_work_mem that
reserves a wodge of shm in the main region, and hand out 'fast DSM
segments' from there, or some other higher level abstraction that's
wired into the resource release system; they would benefit from
huge_pages=try on Linux, they'd be entirely allocated (in the VM
sense) and there'd be no system calls, though admittedly there'd be
more ways for things to go wrong...)