CPU overhead from "Disable concurrency ticket when ACL CACHE lock is acquired in exclusive mode" reduces scan throughput a lot

[mdcallag]

The diff with the title Disable concurrency ticket when ACL CACHE lock is acquired in exclusive mode reduces scan throughput by 32% and 15% on the two servers where I tested it. The issue is new CPU overhead.

Scan throughput is tested via sysbench with a cached database and 1 thread (user, connection, etc).

The servers I used to test it are:

• beelink - Beelink SER 4700u with Ryzen 7 4700u, 16G RAM
• asus - ASUS ExpertCenter PN53 with AMD Ryzen 7 7735HS, 32G RAM

I ran sysbench with 1 table, 30M rows for beelink, 50M rows for asus and the table scan query is done in a loop so that scans run for 5+ minutes.

When I compare builds with and without the diff using git sha b72c7e5 as is (with the diff) and with the new code removed from skip_concurrency_ticket (without the diff) throughput improves by ~32% on beelink and by ~15% on asus. One difference between beelink and asus is that asus has a newer & faster CPU.

The scan query is written like this to avoid finding rows to return, no rows match the WHERE clause:
SELECT * from $table WHERE LENGTH(c) < 0

[mdcallag]

Attached flame graphs show the new CPU overhead in skip_concurrency_ticket and its children. I compare MyRocks 8.0.32 from 24/05/29 at git sha 49b37df (doesn't have the problem) with MyRocks from 24/10/23 at git sha b72c7e5. These flame graphs (*.svg files) are from the beelink server.