« MySQL Central @ OpenWorld 2014 : Speaking about MySQL Performance | Main | My Presentation Slides from MySQL Central @OpenWorld 2014 »

Monday, 29 September, 2014

Indeed, MySQL 5.7 rocks : OLTP_RO Point-Selects 8-tables Benchmark

Indeed, MySQL 5.7 rocks ;-)

This is the part1 of the following blog posts about various benchmark results on MySQL 5.7 - and this particular one is dedicated to the Sysbench OLTP_RO Point-Select 8-tables workload.

We've already published the over 500K QPS on SQL Point-Selects before @32cores-HT server, so may reconfirm it again with MySQL 5.7 DMR5 -vs- other engines :

and now on a similar server, but with 40cores-HT we're able to confirm 645 QPS on the same workload!

If you missed the long story about how we arrived on such a performance level and how to reproduce the test - you may find all here. This workload is the most killing from all Read-Only Sysbench OLTP tests.. And it's really for the first time we started to scale here with MySQL! - look here (all results are from the 40cores-HT server) :

MySQL 5.5 :

MySQL 5.6 :

MySQL 5.7 :

Percona 5.6 :

MariaDB 10.1 :

As you can see, only MySQL 5.7 rocks here ;-)

However, if you remember the full story, this workload is extremely depending on every extra bits and even insignificant overhead.. - personally it's very rare that I saw a RDBMS workload to be so close to HPC problematic ;-)) -- to remind you that to reach 500K QPS (and 645K QPS here) was necessary to use UNIX socket instead of IP socket and an older Sysbench version binary (v0.4.8) which has just little bit less instructions than the v0.4.13 I'm currently use and will be able to get only 575K QPS on this server..

will tell you yet more during my talk on Wednesday, 1/Oct.

to be continued..

Posted by Dimitri at 23:33
Categories: MySQL
blog comments powered by Disqus
Note: if you don't see any "comment" dialog above, try to access this page with another web browser, or google for known issues on your browser and DISQUS..