walredo: simple tests and bench updates

[koivunej]

Separated from #2875.

The microbenchmark has been validated to show similar difference as to larger scale OLTP benchmark.

[koivunej]

validated to show similar difference

In current async-walredo, performance governor, boost disabled, same firefox background noise, trimmed out output:

short/short/1           time:   [23.856 µs 24.087 µs 24.351 µs]
                        change: [+102.34% +104.62% +107.46%] (p = 0.00 < 0.05)
short/short/2           time:   [44.501 µs 45.072 µs 45.649 µs]
                        change: [-8.6098% -7.2657% -5.7848%] (p = 0.00 < 0.05)
short/short/4           time:   [64.947 µs 66.075 µs 67.288 µs]
                        change: [-35.430% -34.125% -32.647%] (p = 0.00 < 0.05)
short/short/8           time:   [91.637 µs 93.185 µs 95.520 µs]
                        change: [-49.495% -48.590% -47.459%] (p = 0.00 < 0.05)
short/short/16          time:   [175.49 µs 176.69 µs 177.94 µs]
                        change: [-56.441% -56.128% -55.795%] (p = 0.00 < 0.05)

medium/medium/1         time:   [214.42 µs 215.18 µs 215.95 µs]
                        change: [+13.991% +14.437% +14.893%] (p = 0.00 < 0.05)
medium/medium/2         time:   [413.60 µs 414.96 µs 416.45 µs]
                        change: [-0.6190% -0.2453% +0.0944%] (p = 0.20 > 0.05)
medium/medium/4         time:   [784.73 µs 786.95 µs 789.17 µs]
                        change: [-4.5931% -4.2810% -3.9914%] (p = 0.00 < 0.05)
medium/medium/8         time:   [859.98 µs 868.79 µs 880.23 µs]
                        change: [-46.889% -46.406% -45.588%] (p = 0.00 < 0.05)
medium/medium/16        time:   [1.0130 ms 1.0229 ms 1.0342 ms]
                        change: [-69.272% -68.951% -68.606%] (p = 0.00 < 0.05)

In OLTP benchmark with pgbench -i -s 10, pgbench -S -c 1 (select-after-insert), pgbench -c 10 (update), pgbench -S -c 1 (select-after-update), the differences were similar-ish:

• x0.82 for select-after-insert
• x1.09 for update
• x0.81 for select-after-update

Setup for this was plain empty neon_local pg start main without any configuration. It seems the /1 (single request) performance matters the most.

[LizardWizzard]

I'm probably not the best fit for reviewing walredo stuff, since I was not heavily involved in recent discussions. So I think someone else needs to have a look. To me looks reasonable

[koivunej]

Thanks! These do work fine as is test cases for #2947 so I am inclined to merge them as is. Asking for another reviewer.

[koivunej]

Should had rebased, doing it now...

[koivunej]

Rebased on top of main.

[koivunej]

Ok, had forgotten this, thought it had already been merged. Let's see what the conflict is.