Tuesday, September 5, 2017

In-memory sysbench and a small server

I repeated the in-memory sysbench test on a small server and the results are similar to what I measured on the large server.

tl;dr
  • MyRocks does worse than InnoDB and the difference is larger on read-heavy tests.
  • InnoDB QPS tends to decrease after 5.6.35
  • InnoDB range scans are more efficient after 5.6.35

Configuration

I used core i5 Intel NUC servers with an SSD, my sysbench helper scripts and sysbench branch. The test was run with 4 tables and 1M rows/table. Somewhat accurate my.cnf files are here. The test was run for MyRocks compiled on August 15 with git hash 0d76ae and InnoDB from upstream 5.6.35, 5.7.17 and 8.0.2. The binlog was enabled but sync on commit was disabled for the binlog and database log. InnoDB 8.0.2 used latin1 charset and latin1_swedish_ci collation to match earlier releases. The sysbench client and mysqld shared the host.

Results

All of the data is here. Below I share the QPS ratios to compare QPS for MyRocks and InnoDB 5.7.17 with InnoDB 5.6.35. The results are similar to the results from in-memory sysbench on a large server.

QPS ratio:
* rocks = QPS for MyRocks / QPS for InnoDB 5.6.35
* inno = QPS for InnoDB 5.7.17 / QPS for InnoDB 5.6.35

1 connection
rocks   inno
0.391   0.966   update-inlist
0.723   0.911   update-one
0.965   1.042   update-index
0.592   0.908   update-nonindex
0.651   0.915   update-nonindex-special
0.700   0.893   delete-only
0.628   1.122   read-write.range100
0.838   1.495   read-write.range10000
0.680   1.155   read-only.range100
0.826   1.438   read-only.range10000
0.655   0.882   point-query
0.247   0.966   random-points
0.287   0.944   hot-points
0.696   0.884   insert-only
-----   -----
0.634   1.037   average

No comments:

Post a Comment