Re: Flushing large data immediately in pqcomm

Hi,

On 2024-04-06 14:34:17 +1300, David Rowley wrote:
> I don't see any issues with v5, so based on the performance numbers
> shown on this thread for the latest patch, it would make sense to push
> it.  The problem is, I just can't recreate the performance numbers.
>
> I've tried both on my AMD 3990x machine and an Apple M2 with a script
> similar to the test.sh from above.  I mostly just stripped out the
> buffer size stuff and adjusted the timing code to something that would
> work with mac.

I think there are a few issues with the test script leading to not seeing a
gain:

1) I think using the textual protocol, with the text datatype, will make it
   harder to spot differences. That's a lot of overhead.

2) Afaict the test is connecting over the unix socket, I think we expect
   bigger wins for tcp

3) Particularly the larger string is bottlenecked due to pglz compression in
   toast.


Where I had noticed the overhead of the current approach badly, was streaming
out basebackups. Which is all binary, of course.


I added WITH BINARY, SET STORAGE EXTERNAL and tested both unix socket and
localhost. I also reduced row counts and iteration counts, because I am
impatient, and I don't think it matters much here. Attached the modified
version.


On a dual xeon Gold 5215, turbo boost disabled, server pinned to one core,
script pinned to another:


unix:

master:
Run 100 100 1000000: 0.058482377
Run 1024 10240 100000: 0.120909810
Run 1024 1048576 2000: 0.153027916
Run 1048576 1048576 1000: 0.154953512

v5:
Run 100 100 1000000: 0.058760126
Run 1024 10240 100000: 0.118831396
Run 1024 1048576 2000: 0.124282503
Run 1048576 1048576 1000: 0.123894962


localhost:

master:
Run 100 100 1000000: 0.067088000
Run 1024 10240 100000: 0.170894273
Run 1024 1048576 2000: 0.230346632
Run 1048576 1048576 1000: 0.230336078

v5:
Run 100 100 1000000: 0.067144036
Run 1024 10240 100000: 0.167950948
Run 1024 1048576 2000: 0.135167027
Run 1048576 1048576 1000: 0.135347867


The perf difference for 1MB via TCP is really impressive.

The small regression for small results is still kinda visible, I haven't yet
tested the patch downthread.

Greetings,

Andres Freund