Re: Improve eviction algorithm in ReorderBuffer

Heikki Linnakangas <hlinnaka@iki.fi>

From: Heikki Linnakangas <hlinnaka@iki.fi>
To: Michael Paquier <michael@paquier.xyz>, Jeff Davis <pgsql@j-davis.com>
Cc: Masahiko Sawada <sawada.mshk@gmail.com>, Amit Kapila <amit.kapila16@gmail.com>, vignesh C <vignesh21@gmail.com>, Peter Smith <smithpb2250@gmail.com>, Tomas Vondra <tomas.vondra@enterprisedb.com>, "Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>, Shubham Khanna <khannashubham1197@gmail.com>, Dilip Kumar <dilipbalaut@gmail.com>, PostgreSQL Hackers <pgsql-hackers@lists.postgresql.org>
Date: 2024-04-10T05:30:22Z
Lists: pgsql-hackers
On 10/04/2024 07:45, Michael Paquier wrote:
> On Tue, Apr 09, 2024 at 09:16:53PM -0700, Jeff Davis wrote:
>> On Wed, 2024-04-10 at 12:13 +0900, Michael Paquier wrote:
>>> Wouldn't the best way forward be to revert
>>> 5bec1d6bc5e3 and revisit the whole in v18?
>>
>> Also consider commits b840508644 and bcb14f4abc.
> 
> Indeed.  These are also linked.

I don't feel the urge to revert this:

- It's not broken as such, we're just discussing better ways to 
implement it. We could also do nothing, and revisit this in v18. The 
only must-fix issue is some compiler warnings IIUC.

- It's a pretty localized change in reorderbuffer.c, so it's not in the 
way of other patches or reverts. Nothing else depends on the binaryheap 
changes yet either.

- It seems straightforward to repeat the performance tests with whatever 
alternative implementations we want to consider.

My #1 choice would be to write a patch to switch the pairing heap, 
performance test that, and revert the binary heap changes.

-- 
Heikki Linnakangas
Neon (https://neon.tech)




Commits

  1. Revert indexed and enlargable binary heap implementation.

  2. Replace binaryheap + index with pairingheap in reorderbuffer.c

  3. Improve eviction algorithm in ReorderBuffer using max-heap for many subtransactions.

  4. Add functions to binaryheap for efficient key removal and update.

  5. Make binaryheap enlargeable.