Table scan performance drops by 29%

The relative QPS is 0.71 for the scan test I run via sysbench.
The relative QPS is: (QPS for my version / QPS for 10.2.30)

My repro script is here, that runs all of my sysbench tests. I ran it with 40 clients, 8 tables and 10M rows per table on a 48-core server.

The changes in relative QPS per release are here. The results for scan have more changes than the other sysbench tests. I see large decreases in:

• 10.5.6 and 10.5.29 but that is fixed in 10.6.5 - see col-5, col-6 and col-7 here
• 10.11.14 - but that is fixed, see col-15 here
• 11.2.6 - but that is fixed, see col-18 here
• 11.7.2 - and this problem remains in 12.1.1, see col-23 to col-26 here

From vmstat I see large increases in CPU overhead and context switch rates starting in 11.7, although there are also large increases after 10.2 and before 11.7. See cpu/o and cs/o here.

Using vmstat metrics for 11.7 relative to 10.2.30 I see large increases in CPU overhead and context switch rates, see cpu/o and cs/o here

Using absolute values from vmstat I see a large increase in the context switch rate on the server from 376 per second in 10.2.30 to 538957 per second in 12.1.1.