Wednesday, June 21, 2017

Linux perf and the CPU regression in MySQL 5.7

Update - the regression isn't as bad as I have been reporting. Read this post to understand why.

I used Linux perf to get more details on system performance while running the point-query test with modern sysbench. This is for an in-memory workload and part of my series on low-concurrency CPU regressions for bug 86215.

  • I have more explaining to do
  • The increase in instructions/query explains the decrease in queries/second from MySQL 5.6 to 5.7 to 8.0.


I tested MySQL using upstream 5.6.35, 5.7.17 and 8.0.1. For 8.0.1 I used the latin1 charset and latin1_swedish_ci collation. My servers are described here and are named i3 NUC and i5 NUC.

My usage of sysbench, including the tests run, and the my.cnf used for each database engine is described here for MySQL 5.6, 5.7 and 8. The my.cnf files I share were for the i3 NUC. For the i5 NUC the InnoDB buffer pool and IO capacity options were increased using these values. The my.cnf files are here for 5.05.1 and 5.5. I use the same server for mysqld and the sysbench clients. The binlog is enabled but sync-on-commit is disabled. Sysbench is run with 4 tables and 1M rows per table.

After loading the tables via the sysbench prepare option I ran the update-index test for a few minutes and then ran point-query with 1 client. After letting point-query warm up for 60 seconds I used this script to collect data from Linux perf. Output from Linux perf is here:


The charts below show the queries/second and instructions/query for each server/release with the point-query sysbench test and 1 client. For the i3 NUC the regression is from MySQL 5.7.17 to 8.0.1. For the i5 NUC it is from 5.6.35 to 5.7.17. Hopefully I will explain why the results are different between the i3 and i5 NUC servers. I am still collecting results from the CPUs I use at work and they don't match what I report here. So I have some explaining to do.

But it is clear to me that the increase in instructions/query explains the decrease in queries/second. Note that I run the sysbench client on the same host as mysqld so the instructions/query overhead includes the sysbench client and the real regression from MySQL 5.6 to 8.0 would be larger were that excluded.


No comments:

Post a Comment