maria-developers team mailing list archive

[Axel Schwenke <axel@xxxxxxxxxxxx>]

Hi,

I have run another sysbench with in-memory workload to check where we are
with respect to mutex contention.

The benchmark is sysbench OLTP (using the bzr trunk of sysbench with LUA
support). 10 mio rows in a single table (about 2.5G .ibd file). InnoDB
buffer pool set to 32G and warmed up.

I have run the test on two different machines. First on our old trusty
facebook2 machine with 8 cores/16 threads in two NUMA zones. Second on the
fat lizard2 machine with 32 cores/64 threads in 4 NUMA zones.

As expected, the heavy contention of the LOCK_open mutex is more visible on
the bigger machine (I have not done any detailed analysis in this benchmark,
but before while doing benchmarks for MDEV-5081)

I notice a visible performance gain in MariaDB-10.0 over MariaDB-5.5 at 32
threads. This is probably due to the work on the InnoDB kernel mutex
(impossible to prove since 5.5 doesn't have mutex instrumentation). At
higher concurrency however the LOCK_open mutex kicks in and performance is
clearly dominated by that mutex alone (it flattens out for all 5.5 and 10.0
versions of MariaDB)

The good news is, that 10.0.4, 10.0.5 and 10.0.6 are quite close. There is a
small regression from 10.0.4 to 10.0.5, but IMHO this isn't worth
investigation while the LOCK_open issue is not fixed.

Attached: diagrams. OpenOffice spreadsheets.


XL

Attachment: sysbench45_8cores.ods
Description: application/vnd.oasis.opendocument.spreadsheet

Attachment: sysbench45_32cores.ods
Description: application/vnd.oasis.opendocument.spreadsheet