Re: Downport of InnoDB/ICP from mysql 5.6 increased Innodb_buffer_pool_read_requests
Ways to repeat:

1. Make the dataset:
./sysbench --test=./tests/db/oltp.lua --oltp-table-size=10000000 \
--db-driver=mysql --mysql-user=root --mysql-socket=/tmp/mysqld.sock \
--mysql-table-engine=innodb prepare

2. Prepare two trees,
1. "Before" tree: 5.3 branched at -r2954
2. "After" tree: 5.3, branched at -r2955, with 3406

Run the attached script and watch the difference in output:

"Before" tree produces (showing only lines with 'OUT'):

incr OUT
13 OUT
incr OUT
12 OUT
incr OUT
12 OUT
incr OUT
12 OUT
incr OUT
12 OUT

"After" tree produces (showing only lines with 'OUT'):
incr OUT
14 OUT
incr OUT
13 OUT
incr OUT
13 OUT
incr OUT
13 OUT
incr OUT
13 OUT

The script.

• it is intended to be run from sql/ directory of a tree
• modify to point to your my.cnf file, which should specify a datadir with pre-loaded sysbench data as specified in earlier comments.
  check-innodb-counters-here.sh
  LPexportBug929838_check-innodb-counters-here.sh

Re: Downport of InnoDB/ICP from mysql 5.6 increased Innodb_buffer_pool_read_requests
The script.

• it is intended to be run from sql/ directory of a tree
• modify to point to your my.cnf file, which should specify a datadir with pre-loaded sysbench data as specified in earlier comments.

If you have trouble compiling the tree because of "warnings being treated as errors", here is a patch that shuts that off.
werror-be-gone.diff
LPexportBug929838_werror-be-gone.diff

Re: Downport of InnoDB/ICP from mysql 5.6 increased Innodb_buffer_pool_read_requests
If you have trouble compiling the tree because of "warnings being treated as errors", here is a patch that shuts that off.

Re: Downport of InnoDB/ICP from mysql 5.6 increased Innodb_buffer_pool_read_requests
The dataset is at:

work.askmonty.org:/tmp/sysbench-data-10m.sql.bz2

Re: Downport of InnoDB/ICP from mysql 5.6 increased Innodb_buffer_pool_read_requests
We have a performance degradation for sysbench that was explained in Olav's Sandstaa patch for mysql-5.6. (see rev 3639.1.70 of lp:mysql-server/trunk )
Olav dis not suggest any more or less general solution. He just prohibited using ICP for innodb primary clustered keys.

Launchpad bug id: 929838

The problem doesn't seem to be introduced in MariaDB. It is innodb+xtradb changes. I'm still keeping this open, because I intend to investigate this and see if there are some redundant reads done as a result.