On 2023/9/5 16:15, John Naylor wrote:
On Thu, Aug 24, 2023 at 5:55 PM Quan Zongliang <quanzongli...@yeah.net
<mailto:quanzongli...@yeah.net>> wrote:
> In the function heapgetpage. If a table is not updated very frequently.
> Many actions in tuple loops are superfluous. For all_visible pages,
> loctup does not need to be assigned, nor does the "valid" variable.
> CheckForSerializableConflictOutNeeded from
> HeapCheckForSerializableConflictOut function, it only need to inspect at
Thanks for submitting! A few weeks before this, there was another
proposal, which specializes code for all paths, not just one. That patch
also does so without duplicating the loop:
https://www.postgresql.org/message-id/20230716015656.xjvemfbp5fysj...@awork3.anarazel.de
<https://www.postgresql.org/message-id/20230716015656.xjvemfbp5fysj...@awork3.anarazel.de>
Nice patch. I'm sorry I didn't notice it before.
> the beginning of the cycle only once. Using vtune you can clearly see
> the result (attached heapgetpage.jpg).
>
> So by splitting the loop logic into two parts, the vtune results show
> significant improvement (attached heapgetpage-allvis.jpg).
For future reference, it's not clear at all from the screenshots what
the improvement will be for the user. In the above thread, the author
shares testing methodology as well as timing measurements. This is
useful for reproducibilty, as well as convincing others that the change
is important.
Here's how I test it
EXPLAIN ANALYZE SELECT * FROM orders;
Maybe the test wasn't good enough. Although the modified optimal result
looks good. Because it fluctuates a lot. It's hard to compare. The
results of vtune are therefore used.
My patch is mainly to eliminate:
1, Assignment of "loctup" struct variable (in vtune you can see that
these 4 lines have a significant overhead: 0.4 1.0 0.2 0.4).
2. Assignment of the "valid" variable.(overhead 0.6)
3. HeapCheckForSerializableConflictOut function call.(overhead 0.6)
Although these are not the same overhead from test to test. But all are
too obvious to ignore. The screenshots are mainly to show the three
improvements mentioned above.
I'll also try Andres Freund's test method next.
--
John Naylor
EDB: http://www.enterprisedb.com <http://www.enterprisedb.com>
