TagRenamingFilter

[etki]

This is the last of performance-related PRs. As with previous one, this description will be updated once i have more time to run all the benchmarks. Nothing super super cool, done mostly to keep consistency, but these are the latest results (requires additional verification). Please keep in mind that X is not X supplied tags in every input, but only in 90% of them (other 10% contain 1-64 tags, done to simulate a somewhat real situation).

processed tags	version	nanoseconds / operation	instructions
0	main	91.356 ± 0.353	374.790
0	pr	81.980 ± 0.859	335.794
1	main	126.592 ± 0.345	473.020
1	pr	92.874 ± 0.277	350.152
4	main	146.673 ± 0.369	568.136
4	pr	106.541 ± 0.284	409.278
8	main	185.319 ± 0.457	725.593
8	pr	127.259 ± 0.235	500.279
16	main	274.408 ± 0.560	1162.592
16	pr	191.191 ± 0.331	800.652
32	main	503.440 ± 0.745	2126.366
32	pr	395.133 ± 0.781	1817.026
64	main	821.375 ± 1.231	4549.852
64	pr	865.062 ± 1.071	5404.027

The results are getting worse in 32+ range (compared to the single-iteration version as well), because the chance of actually seeing a replacement gets to 50% and higher. This, i believe, is where we are actually leaving the real situation. The double-iteration version is in fact the last minute experiment, there might be a bit of an improvement, will hopefully have some time to investigate. As in all other PRs, #6113 would greatly help to make another CPU relief step.

[jonatan-ivanov]

Hi,

Instead of replying to all of your PRs separately, let me reply here to all of them first.
First of all: thank you very much!

I looked through them quickly, here are a few updates/things to consider:

• We need to discuss them internally (they are somewhat connected)
• Timeline-wise we can consider them for 1.16.0 (which means the earliest we can merge them is mid next month)
• I know @shakuzen wanted to create JMH benchmarks that are connected to these areas, I think we should do that first
• Last week I worked on making it easier to run perf tests in a more stable infrastructure than our laptops, I need to discuss this with @shakuzen, it can help merging your PRs if we can use it to run stable JMH benchmarks

We will get back to you soon with more update after we discussed the above/reviewed your PRs.

[etki]

We need to discuss them internally (they are somewhat connected)

Absolutely no problem. I'm just having fun here, feel free to reject what you don't see as fitting. I've separate all of them so that can be done easily without hassle. This also gives me some time to get back and properly describe everything that is done here.

I know @shakuzen wanted to create JMH benchmarks that are connected to these areas, I think we should do that first

I have some in #6174 - if there was no work done previously, then you might want to take a look at them. They use existing API, so it should be possible to throw it at any recent commit.

I need to discuss this with @shakuzen, it can help merging your PRs if we can use it to run stable JMH benchmarks

I also hope that i'll get my homelab to a better place soon, i have more than one minibox with different CPUs. Just to reiterate, all of these were done on a pristine Fedora (i.e. no additional services running) and intel N100 with fixed frequency, they were quite stable with <1% fluctuation.