Re: BitmapHeapScan streaming read user and prelim refactoring

On 2/14/25 18:50, Tomas Vondra wrote:
> On 2/14/25 18:31, Melanie Plageman wrote:
>> io_combine_limit 1, effective_io_concurrency 16, read ahead kb 16
>>
>> On Fri, Feb 14, 2025 at 12:18 PM Tomas Vondra <tomas(at)vondra(dot)me> wrote:
>>>
>>> Based on off-list discussion with Melanie, I ran a modified version of
>>> the benchmark, with these changes:
>>
>> Thanks! It looks like the worst offender is io_combine_limit 1 (128
>> kB), effective_io_concurrency 16, read_ahead_kb 16. This is a
>> combination that people shouldn't run in practice I think -- an
>> io_combine_limit larger than read ahead.
>>
>> But we do still see regressions with io_combine_limit 1,
>> effective_io_concurrency 16 at other read_ahead_kb values. Therefore
>> I'd be interested to see that subset (ioc 1, eic 16, diff read ahead
>> values) with the attached patch.
>>
>>> FWIW this does not change anything in the detection of sequential access
>>> patterns, discussed nearby, because the benchmarks started before Andres
>>> looked into that. If needed, I can easily rerun these tests, I just need
>>> a patch to apply.
>>
>> Attached a patch that has all the commits squashed + removes
>> sequential detection.
>>
>
> OK, it's running.
>
> I didn't limit the benchmarks any further, it seems useful to have
> "full" comparison with the last run. I expect to have the results in ~48
> hours, i.e. by Monday.
>

OK, I've uploaded the results to the github repository as usual

https://github.com/tvondra/bitmapscan-tests/tree/main/20250214-184807

and I've generated the same PDF reports, with the colored comparison.

If you compare the pivot tables (I opened the "same" PDF from the two
runs and flip between them using alt-tab, which makes the interesting
regions easy to spot), the change is very clear.

Disabling the sequential detection greatly reduces the scope of
regressions. That looks pretty great, IMO.

It also seems to lose some speedups, especially with io_combine_limit=1
and eic=1. I'm not sure why, if that's expected, etc.

There still remain areas of regression, but most of them are for cases
that'd use index scan (tiny fraction of rows scanned), or with
read-ahead=4096 (and not for the lower settings).

The read-ahead dependence is actually somewhat interesting, because I
realized the RAID array has this set to 8192 by default, i.e. even
higher than 4096 where it regresses. I suppose mdadm does that, or
something, I don't know how the default is calculated. But I assume it
depends on the number of devices, so larger arrays might have even
higher read-ahead values.

I'm running the benchmarks with ra=8192, shouldn't take more than a
couple hours.

regards

--
Tomas Vondra