Monday, December 11, 2017

Sysbench: in-memory and a fast server, part 2

This post has results for in-memory sysbench with 1 table and 8M rows/table. My previous post was for in-memory sysbench with 8 tables and 1M rows/table. The goal is to understand whether performance is lost when there is more contention for a table, and in some cases more contention for the same rows. This repeats tests I ran in September and the results are similar.

There are four tests for which QPS with 1 table is much worse than with 8 tables:
  • update-one - all engines do worse. This is expected.
  • random-points - InnoDB and TokuDB do worse
  • hot-points - all engines do worse
  • insert - InnoDB and TokuDB do worse

All of the data is on github. The graphs below have the absolute QPS for 48 concurrent clients.

update-one

All engines do worse with 1 table. I am not surprised because only 1 row gets all updates while the updates are spread across 8 rows in the 8-table test.

random-points

InnoDB and TokuDB do worse with 1 table. My previous post shows where the contention occurs.


hot-points

All engines do worse with 1 table. My previous post shows where the contention occurs. Issue 674 is open to make this better in MyRocks.

insert

InnoDB and TokuDB do worse with 1 table. I have yet to debug this.



No comments:

Post a Comment