Affects Version/s: None
Fix Version/s: None
**Reported in Launchpad by Axel Schwenke last update 30-03-2017 03:03:59
The fix for lp:1433432 in Percona Server 5.6 introduces a severe performance regression. Sysbench OLTP read-only throughput drops by 50% (single thread) up to 95% (40 threads on a 16-core machine).
In order to see that, the active set of the database must be slightly bigger than the buffer pool and the workload must access rows in a rather random fashion.
Here are some numbers. Command line:
sysbench-0.4.12 --test=oltp --oltp-read-only --oltp-range-size=1000 --oltp-dist-type=uniform --oltp-table-size=1000000 --mysql-socket=/tmp/mysqld.sock.sysbench --mysql-user=root --num-threads=40 --max-time=30 run
5.6.27: 660 tps
5.6.28: 36 tps
This doesn't change when using multiple tables. Performance comes back to normal when the buffer pool is increased. I think this performance regression is too big and too likely to being hit by a real-world user to be ignored. It is quite common that the active set of a database does not fit into the buffer pool.
Suggestion: roll back commit 6532572a783ea5a025561eda9b1a9350c73405a6