There are a few things up in the air, so I'm coming back to this list to summarize and add a recent update:
On Mon, Nov 14, 2022 at 7:59 PM John Naylor <
john.naylor@enterprisedb.com> wrote:
>
> - See how much performance we actually gain from tagging the node kind.
Needs a benchmark that has enough branch mispredicts and L2/3 misses to show a benefit. Otherwise either neutral or worse in its current form, depending on compiler(?). Put off for later.
> - Try additional size classes while keeping the node kinds to only four.
This is relatively simple and effective. If only one additional size class (total 5) is coded as a placeholder, I imagine it will be easier to rebase shared memory logic than using this technique everywhere possible.
> - Optimize node128 insert.
I've attached a rough start at this. The basic idea is borrowed from our bitmapset nodes, so we can iterate over and operate on word-sized (32- or 64-bit) types at a time, rather than bytes. To make this easier, I've moved some of the lower-level macros and types from bitmapset.h/.c to pg_bitutils.h. That's probably going to need a separate email thread to resolve the coding style clash this causes, so that can be put off for later. This is not meant to be included in the next patchset. For demonstration purposes, I get these results with a function that repeatedly deletes the last value from a mostly-full node128 leaf and re-inserts it:
select * from bench_node128_load(120);
v11
NOTICE: num_keys = 14400, height = 1, n1 = 0, n4 = 0, n15 = 0, n32 = 0, n61 = 0, n128 = 121, n256 = 0
fanout | nkeys | rt_mem_allocated | rt_sparseload_ms
--------+-------+------------------+------------------
120 | 14400 | 208304 | 56
v11 + 0006 addendum
NOTICE: num_keys = 14400, height = 1, n1 = 0, n4 = 0, n15 = 0, n32 = 0, n61 = 0, n128 = 121, n256 = 0
fanout | nkeys | rt_mem_allocated | rt_sparseload_ms
--------+-------+------------------+------------------
120 | 14400 | 208816 | 34
I didn't test inner nodes, but I imagine the difference is bigger. This bitmap style should also be used for the node256-leaf isset array simply to be consistent and avoid needing single-use macros, but that has not been done yet. It won't make a difference for performance because there is no iteration there.
> - Try templating out the differences between local and shared memory.
I hope to start this sometime after the crashes on 32-bit are resolved.
--
John Naylor
EDB:
http://www.enterprisedb.com
