Why MySQL 5.6–5.7 Shows the Biggest QPS Drop in Low‑Concurrency Tests

Background

After publishing benchmark results for MySQL 5.6, 5.7 and 8, the author extended the study to older releases (5.0, 5.1, 5.5) to complete a low‑concurrency performance series.

The author notes that MySQL 4.1 and 5.5 performed poorly and were omitted from the main comparison, and that the most significant QPS decline occurs between 5.6 and 5.7, often exceeding the drop between 5.0 and 5.6. The underlying cause is referenced in Bug 86215.

Configuration

Tests were run with upstream MySQL versions 5.0.96, 5.1.72, 5.5.51, 5.6.35, 5.7.17 and 8.0.1 on an i5 NUC. MySQL 8.0.1 used the latin1 charset and latin1_swedish_ci collation.

MySQL 4.1.22 was compiled on the same hardware but not reported because of poor results; MySQL 4.0 compiled on Ubuntu 16.04 with gcc 4.7/4.8 segfaults shortly after start‑up.

Sysbench was used to drive the workload. The my.cnf files for each version (especially 5.0, 5.1, 5.5) are provided in the original links. For the i5 NUC the InnoDB buffer‑pool size and IO‑capacity options were increased. Binlog was enabled while sync‑on‑commit was disabled. Four tables of 1 M rows each were used, and tests were executed with 1, 2 and 4 client threads, storing data in the InnoDB buffer pool.

Results

All QPS measurements are recorded at the provided link, and selected charts are shown below.

The table compares each version’s QPS to MySQL 5.0. For example, a value of 0.53 for MySQL 8 (update‑index) means MySQL 8 achieves only 53 % of MySQL 5.0’s QPS, i.e., MySQL 5.0 is roughly twice as fast. The drop from 5.6 to 5.7 is dramatic, while the transition from 5.7 to 8 does not repeat the decline.

Charts

For the update‑index test the largest QPS drop occurs between 5.6 and 5.7.

For the update‑nonindex test the biggest drop is also between 5.6 and 5.7; a notable decline also appears between 5.1 and 5.5, which was already resolved in 5.6. Version 5.5 may be considered a poor release.

For read‑write.range100 and read‑write.range10000 the maximum QPS reduction again appears between 5.6 and 5.7.

For read‑only.range10 and read‑only.range10000 the same pattern holds.

For the point‑query test the biggest QPS drop is also between 5.6 and 5.7.

The insert test shows only a slight QPS decline across versions.

Conclusion

The study confirms that the most pronounced QPS reduction in low‑concurrency scenarios occurs between MySQL 5.6 and 5.7, while the jump to MySQL 8 does not repeat the drop. The author originally expected a steady performance decline with newer versions, but the data suggests most regressions are isolated to 5.7 and may be addressable.

These findings reflect the author’s perspective; readers are encouraged to discuss alternative results with the original author.