Re: Generating partitioning tuple conversion maps faster

On 29 June 2018 at 11:10, David Rowley <david(dot)rowley(at)2ndquadrant(dot)com> wrote:
> On further inspection, the slowdown is coming from the planner in
> make_inh_translation_list(). The INSERT path does not hit that since
> the planner's job is pretty simple for simple INSERTs.

I've attached a patch that uses SearchSysCacheAttName to speed up
these translations in the planner.

This puts test 6 more at the level I'd have expected.

Here are fresh benchmarks results taken with the attached, again on
AWS m5d instance, though probably not the same one as before
(fsync=off).

Unpatched:

Test 1
tps = 922.479156 (excluding connections establishing)

Test 2
tps = 334.701555 (excluding connections establishing)

Test 3
tps = 327.547316 (excluding connections establishing)

Test 4
tps = 1198.924131 (excluding connections establishing)

Test 5
tps = 125.130723 (excluding connections establishing)

Test 6
tps = 81.511072 (excluding connections establishing)

Patched

Test 1
tps = 918.105382 (excluding connections establishing)

Test 2
tps = 913.315387 (excluding connections establishing)

Test 3
tps = 893.578988 (excluding connections establishing)

Test 4
tps = 1213.238177 (excluding connections establishing)

Test 5
tps = 459.022550 (excluding connections establishing)

Test 6
tps = 416.835747 (excluding connections establishing)

--
David Rowley http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services