Horton

Re: BitmapHeapScan streaming read user and prelim refactoring

Tomas Vondra <tomas@vondra.me>

View thread View on postgr.es
From: Tomas Vondra <tomas@vondra.me>
To: Robert Haas <robertmhaas@gmail.com>, Melanie Plageman <melanieplageman@gmail.com>
Cc: Nazir Bilal Yavuz <byavuz81@gmail.com>, Dilip Kumar <dilipbalaut@gmail.com>, Heikki Linnakangas <hlinnaka@iki.fi>, Thomas Munro <thomas.munro@gmail.com>, Andres Freund <andres@anarazel.de>, Pg Hackers <pgsql-hackers@postgresql.org>
Date: 2025-02-12T20:07:51Z
Lists: pgsql-hackers 

On 2/12/25 20:08, Robert Haas wrote:
> On Mon, Feb 10, 2025 at 4:41 PM Melanie Plageman
> <melanieplageman@gmail.com> wrote:
>> I had taken to thinking of it as "queue depth". But I think that's not
>> really accurate. Do you know why we set it to 1 as the default? I
>> thought it was because the default should be just prefetch one block
>> ahead. But if it was meant to be spindles, was it that most people had
>> one spindle (I don't really know what a spindle is)?
> 
> AFAIK, we're imagining that people are running PostgreSQL on a hard
> drive like the one my desktop machine had when I was in college, which
> indeed only had one spindle. There were one or possibly several
> magnetic platters rotating around a single point. But today, nobody
> has that, certainly not on a production machine. For instance if you
> are using RAID or LVM you have several drives hooked up together, so
> that would be multiple spindles even if all of them were HDDs, but
> probably they are all SSDs which don't have spindles anyway. If you
> are on a virtual machine you have a slice of the underlying machine's
> I/O and whatever that is it's probably something more complicated than
> a single spindle.

I think we've abandoned the idea that effective_io_concurrency has
anything to do with the number of spindles - it was always a bit wrong
because even spinning rust had TCQ/NCQ (i.e. ability to optimize head
movement for a queue of requests). But it completely lost the meaning
with flash storage. And in PG13 we even stopped "recalculating" the
prefetch distance, and use raw effective_io_concurrency.

> 
> To be fair, I guess the concept of effective_io_concurrency isn't
> completely meaningless if you think about it as the number of
> simultaneous I/O requests that can be in flight at a time. But I still
> think it's true that it's hard to make any reasonable guess as to how
> you should set this value to get good performance. Maybe people who
> work as consultants and configure lots of systems have good intuition
> here, but anybody else is probably either not going to know about this
> parameter or flail around trying to figure out how to set it.
> 

Yes, the "queue depth" interpretation seems to be much more accurate
than "number of spindles".

> It might be interesting to try some experiments on a real small VM
> from some cloud provider and see what value is optimal there these
> days. Then we could set the default to that value or perhaps somewhat
> more.
> 

AFAICS the "1" value is simply one of the many "defensive" defaults in
our sample config. It's much more likely to help than cause harm, even
on smaller/older systems, but for many systems a higher value would be
more appropriate. There's usually a huge initial benefit (say, going to
16 or 32), and then the benefits diminish fairly quickly.


regards

-- 
Tomas Vondra

Commits

  1. Fix bitmapheapscan incorrect recheck of NULL tuples

  2. Increase default maintenance_io_concurrency to 16

  3. Separate TBM[Shared|Private]Iterator and TBMIterateResult

  4. Improve read_stream.c advice for dense streams.

  5. Increase default effective_io_concurrency to 16

  6. Delay extraction of TIDBitmap per page offsets

  7. Add lossy indicator to TBMIterateResult

  8. Move BitmapTableScan per-scan setup into a helper

  9. Add and use BitmapHeapScanDescData struct

  10. Fix bitmap table scan crash on iterator release

  11. Bitmap Table Scans use unified TBMIterator

  12. Add common interface for TBMIterators

  13. Make table_scan_bitmap_next_block() async-friendly

  14. Move EXPLAIN counter increment to heapam_scan_bitmap_next_block

  15. Refactor tidstore.c iterator buffering.

  16. BitmapHeapScan: Remove incorrect assert and reset field

  17. Change BitmapAdjustPrefetchIterator to accept BlockNumber

  18. BitmapHeapScan: Use correct recheck flag for skip_fetch

  19. BitmapHeapScan: Push skip_fetch optimization into table AM

  20. BitmapHeapScan: postpone setting can_skip_fetch

  21. BitmapHeapScan: begin scan after bitmap creation

  22. Fix EXPLAIN Bitmap heap scan to count pages with no visible tuples

  23. Remove redundant snapshot copying from parallel leader to workers

  24. Remove some obsolete smgrcloseall() calls.

  25. Remove the "snapshot too old" feature.

  26. Compute XID horizon for page level index vacuum on primary.