Horton

Re: [PoC] Improve dead tuple storage for lazy vacuum

John Naylor <john.naylor@enterprisedb.com>

View thread View on postgr.es
From: John Naylor <john.naylor@enterprisedb.com>
To: Masahiko Sawada <sawada.mshk@gmail.com>
Cc: Nathan Bossart <nathandbossart@gmail.com>, Andres Freund <andres@anarazel.de>, Matthias van de Meent <boekewurm+postgres@gmail.com>, Yura Sokolov <y.sokolov@postgrespro.ru>, PostgreSQL-development <pgsql-hackers@postgresql.org>
Date: 2022-10-05T09:40:31Z
Lists: pgsql-hackers 
On Wed, Oct 5, 2022 at 1:46 PM Masahiko Sawada <sawada.mshk@gmail.com>
wrote:
>
> On Wed, Sep 28, 2022 at 12:49 PM Masahiko Sawada <sawada.mshk@gmail.com>
wrote:
> >
> > On Fri, Sep 23, 2022 at 12:11 AM John Naylor
> > <john.naylor@enterprisedb.com> wrote:
> > Yeah, node31 and node256 are bloated.  We probably could use slab for
> > node256 independently. It's worth trying a benchmark to see how it
> > affects the performance and the tree size.

This wasn't the focus of your current email, but while experimenting with
v6 I had another thought about local allocation: If we use the default slab
block size of 8192 bytes, then only 3 chunks of size 2088 can fit, right?
If so, since aset and DSA also waste at least a few hundred bytes, we could
store a useless 256-byte slot array within node256. That way, node128 and
node256 share the same start of pointers/values array, so there would be
one less branch for getting that address. In v6, rt_node_get_values and
rt_node_get_children are not inlined (asde: gcc uses a jump table for 5
kinds but not for 4), but possibly should be, and the smaller the better.

> Regarding DSA support, IIUC we need to use dsa_pointer in inner nodes
> to point to its child nodes, instead of C pointers (ig, backend-local
> address). I'm thinking of a straightforward approach as the first
> step; inner nodes have a union of rt_node* and dsa_pointer and we
> choose either one based on whether the radix tree is shared or not. We
> allocate and free the shared memory for individual nodes by
> dsa_allocate() and dsa_free(), respectively. Therefore we need to get
> a C pointer from dsa_pointer by using dsa_get_address() while
> descending the tree. I'm a bit concerned that calling
> dsa_get_address() for every descent could be performance overhead but
> I'm going to measure it anyway.

Are dsa pointers aligned the same as pointers to locally allocated memory?
Meaning, is the offset portion always a multiple of 4 (or 8)? It seems that
way from a glance, but I can't say for sure. If the lower 2 bits of a DSA
pointer are never set, we can tag them the same way as a regular pointer.
That same technique could help hide the latency of converting the pointer,
by the same way it would hide the latency of loading parts of a node into
CPU registers.

One concern is, handling both local and dsa cases in the same code requires
more (predictable) branches and reduces code density. That might be a
reason in favor of templating to handle each case in its own translation
unit. But that might be overkill.
--
John Naylor
EDB: http://www.enterprisedb.com

Commits

  1. radixtree: Fix SIGSEGV at update of embeddable value to non-embeddable.

  2. Get rid of anonymous struct

  3. Teach radix tree to embed values at runtime

  4. Teach TID store to skip bitmap for small numbers of offsets

  5. Use bump context for TID bitmaps stored by vacuum

  6. Fix alignment of stack variable

  7. Use TidStore for dead tuple TIDs storage during lazy vacuum.

  8. Rethink create and attach APIs of shared TidStore.

  9. Fix inconsistent function prototypes with function definitions.

  10. Fix a calculation in TidStoreCreate().

  11. Fix potential integer handling issue in radixtree.h.

  12. Add TIDStore, to store sets of TIDs (ItemPointerData) efficiently.

  13. Fix link error for test_radixtree module on Windows

  14. Blind attempt to fix ODR violations

  15. Fix incorrect format specifier for int64

  16. Fix redefinition of typedefs

  17. Add template for adaptive radix tree

  18. Fix signedness error in 9f225e992 for gcc

  19. Introduce helper SIMD functions for small byte arrays

  20. Optimize vacuuming of relations with no indexes.

  21. Add bound check before bsearch() for performance

  22. Allocate consecutive blocks during parallel seqscans