Insert benchmark: MySQL 5.6, 5.7 & 8.0 on a small server

I have another performance report for the insert benchmark using MySQL 5.6, 5.7 and 8.0 on a small server. As normal, I had to repeat the tests a few times: I make mistakes and also want results for new my.cnf settings. The goal for the tests is to understand how low-concurrency performance changes in new MySQL releases. Low-concurrency tests are great for finding CPU regressions. High-concurrency tests are also great but I only have small servers at home.

For the CPU-bound setup, new MySQL (8.0) gets between 10% and 37% less throughput than old MySQL (5.6) because 8.0 uses more CPU/query. The increase in CPU/query for MySQL from 5.6 to 8.0 is large while PostgreSQL avoids CPU regressions from 12.4 to 14.0.

Disclaimers:

• Be careful about interpreting these results, the context here is a workload with simple queries.
• These results have yet to be reviewed by others. There can be mistakes.
• 8.0.20 and 8.0.22 were exciting releases. Performance is slowly improving since then.

Details

Tests were run for MySQL versions 5.6.49, 5.7.35, 8.0.20, 8.0.22, 8.0.23, 8.0.26 and 8.0.27. I try to use similar my.cnf settings except when new versions have new features. The options are in the my56, my57 and my80 directories here (look for my.cnf.cy8). Scripts to build MySQL from source are here for MySQL 5.6, 5.7 and 8.0. One of the big changes I made was to add innodb_log_writer_threads=OFF to the MySQL 8.0 my.cnf settings. Without that the CPU overhead was much worse.

The test servers have 4 cores and tests were run with either 1 insert connection or 2 connections (rate limited inserts, not rate-limited queries). This page has links to explain the insert benchmark and perf report format. The test was run in CPU-bound and IO-bound modes. For each mode, X rows were inserted into a table without a secondary index, then 3 secondary indexes were created, then Y more rows were inserted. The values for X and Y were (20M, 20M) for CPU-bound and (500M, 10M) for IO-bound. After the inserts there were three insert+query steps where the insert connection was rate-limited to 100, 500 and 1000 inserts/s. Each query step was run for 2 hours.

The test server uses Ubuntu 20.04, gcc-9.3.0-17 and uname -a shows 5.4.0-89-generic. The database filesystem uses XFS with discard enabled. 

Results

I have shell scripts to generate perf reports. The reports for CPU-bound and IO-bound are here and here. Below I use throughput ratios where the denominator is the value for MySQL 5.6.49 and the numerator is the value for another version. A value less than 1.0 means the new version is slower than 5.6.49.

Comments on CPU-bound (see the summary):

• for the l.i0 test (inserts without secondary indexes) the throughput ratio is 0.63 for 8.0.27. The slowdown is explained by an increase in CPU/query. See the cpupq column here that is 15 for 5.6.49 and 21 for 8.0.27.
• for the l.i1 test (inserts with secondary indexes) the throughput ratio is 0.73 for 8.0.27. The slowdown is explained by an increase in CPU/query. See the cpupq column here that is 32 for 5.6.49 and 43 for 8.0.27.
• for the q100.1, q500.1 and q1000.1 tests the throughput ratios using the qps columns are 0.81, 0.76 and 0.74 for 8.0.27. This is probably explained by more CPU/query. See the cpupq columns for q100.1, q500.1 and q1000.1. The throughput ratios were much better for 8.0.22 than for 8.0.27 (0.94, 0.89 and 0.86) but I have not attempted to debug that. I was unable to explain why the wkbpi column is larger for 8.0.27 than 5.6.49 and larger means more write-amp (KB written/insert). It was larger for 8.0.27 when measured by iostat, but smaller when measured by global status counters like Innodb_dblwr_pages_written, Innodb_os_log_written, Innodb_data_written and Innodb_buffer_pool_pages_flushed. This remains a mystery. I am not sure it is possible to configure InnoDB writeback in 8.0 to match 5.6. The only difference in the my.cnf settings is the use of innodb_idle_flush_pct=1 for the 8.0 my.cnfs.

Comments on IO-bound (see the summary):

• for the l.i0 test (inserts without secondary indexes) results are similar to the CPU-bound results above because this step is still CPU-bound, 8.0 is slower because it uses more CPU/query. See here.
• for the l.x test (create index) 8.0 is faster than 5.6. I think significant changes were done for InnoDB create index. See here.
• for the l.i1 test (inserts with secondary indexes) results for 8.0.26 & 8.0.27 are better than 5.6.49, but results for 8.0.20 and 8.0.22 and 8.0.23 is in between. With the new release process there can be significant changes for performance between point releases. See the ips column here. I think this is related to the big changes to redo log code in 8.0 as CPU/query (cpupq) and context switches/query (cspq) were larger in 8.0.20 and 8.0.22.
• for the q100.1, q500.1 and q1000.1 tests the throughput ratios are larger than 1 as 8.0 is faster than 5.6, except for 8.0.20 which appears to be an exciting release. See the qps and cpupq columns here. MySQL 8.0.27 uses less CPU/query than 5.6.49.

Insert benchmark: PostgreSQL 12, 13 and 14 on a small server

I have another performance report for the insert benchmark using PostgreSQL 12.4, 13.4 and 14.0 on a small server. As normal, I had to repeat the tests a few times: I make mistakes and sometimes need results for new tuning options. The goal for the tests is to understand how low-concurrency performance changes in new PostgreSQL releases. Low-concurrency tests are great for finding CPU regressions. High-concurrency tests are also great but I only have small servers at home.

Executive summary:

• Postgres is boring. There were no regressions.

Details

Tests were run for Postgres versions 12.4, 13.4 and 14.0. I used the same configuration file (see here). The script to build Postgres from source is here.

The test servers have 4 cores and tests were run with either 1 insert connection or 2 connections (rate limited inserts, not rate-limited queries). This page has links to explain the insert benchmark and perf report format. The test was run in CPU-bound and IO-bound modes. For each mode, X rows were inserted into a table without a secondary index, then 3 secondary indexes were created, then Y more rows were inserted. The values for X and Y were (20M, 20M) for CPU-bound and (500M, 10M) for IO-bound. After the inserts there were three insert+query steps where the insert connection was rate-limited to 100, 500 and 1000 inserts/s. Each query step was run for 2 hours.

The test server uses Ubuntu 20.04, gcc-9.3.0-17 and uname -a shows 5.4.0-89-generic. The database filesystem uses XFS with discard enabled. 

Results

I have shell scripts to generate perf reports. The reports for CPU-bound and IO-bound are here and here. From the CPU-bound (see the summary) and IO-bound results (see the summary):

• Postgres is boring. There were no regressions.

Welll actually, LSM and B-tree

 Things I read about LSM that make me go well actually:

• LSM does sequential writes - well actually, writes are sequential logically (per-file) but not physically (per-device). From the physical perspective compaction writes are large & concurrent rather than sequential. Compaction writes to files sequentially but it is usually concurrent (multi-threaded) with a file written sequentially per-thread. There can also be small writes to the WAL. With concurrent compaction threads the large (maybe 1MB+) writes are interleaved at the device.
• Compaction does merge-sort - well actually, that is a k-way merge, not a merge-sort.
• Leveled compaction suffers from write-amplification - well actually, write-amplification was 10X smaller with MyRocks when the first workload was moved to it from InnodDB. Write-amp with leveled compaction is larger than with tiered, but that is the (C)RUM Conjecture in action: leveled means less space-amp at the cost of more write-amp. One of the problems is that the naive (hand-wavy) estimate of write-amp for leveled (~8 or ~10 per level) is frequently too pessimistic.
• LSM suffers from write stalls - well actually, that is more true than the previous bullet points. But the real problem for an LSM is figuring out how to smooth the write stalls (reduce response time variance) and RocksDB needs to get better at that. Everything (b-tree, LSM, etc) suffers from write-stalls when ingest exceeds write back throughput and the easy way to reproduce that is a benchmark with WAL fsync disabled. A b-tree benefits from feedback that serves as a small write-stall - when the buffer pool is full of dirty pages and the working set is not cached then a write likely needs to wait for 1 page to be written back before it can read something into the buffer pool. But an LSM doesn't have that feedback as a write usually just needs something to be put into the memtable, non-unique secondary index maintenance doesn't require reads, and even reads needed to check unique constraints (likely for a PK) can be avoided in some cases with bloom filters. Checkpoint is one place where a b-tree can suffer from large write stalls. Not that many years ago the insert benchmark was great at documenting longer write-stalls related to checkpoint in InnoDB and WiredTiger. Fortunately those have been fixed. Postgres also had serious problems, now fixed, from checkpoint IO creating bursts of write IO that starve other IO (like reads for user queries). Checkpoint is a hard problem.

DeWitt Clause vs the public cloud

The DeWitt Clause is the name of a clause in the terms of service that prevents a customer from disclosing performance of a licensed product. Things turned out OK for David DeWitt and he even went on to work for at least one company that now uses a DeWitt Clause for their database products.

I am not a lawyer but I have a strong opinion on the DeWitt Clause and non-compete agreements. I think both are anti-competitive and hope more is done to limit their usage.

One of the arguments for DeWitt Clauses is that benchmarking is hard, others will get it wrong and that can mislead customers. While the first parts are correct (it is hard, people will get it wrong) the last clause feels like a bad faith argument to me. Vendors haven't always pursued truth while running their benchmarks. I agree there will be plenty of benchmarketing in a world without DeWitt Clauses. I think the benefits outweigh the problems.

Database vendors

We used to think of this as a thing for legacy enterprise database vendors and many of the new vendors don't use it -- not MongoDB Atlas as of 24-Jun-2021, not CockroachDB as of 19-Oct-2021, not TimescaleDB as of 24-Sep-2020 and 5-Mar-2019, not Altinity as of 1-Jul-2021.

But other new vendors use it and this became news after Databricks published a great result for 100TB TPC-DS and a less than great result for SnowflakeDB. They then changed their license to the DeWitt Embrace Clause which is like copyleft for the DeWitt Clause. This means that if you disclose benchmark results for them, they are allowed to do the same for you. I hope the DeWitt Embrace Clause is embraced by more vendors who use a DeWitt Clause.

For the record, Snowflake quickly replied to the Databricks result. Frankly, I prefer to see a benchmarketing battle between Snowflake and Oracle. But it is awesome that the Snowflake response explains that anyone can repeat their result by signing up, spending about $300 and clicking a few buttons. Cloud DBMS is fun. The Snowflake ToS (as of 1-Nov-2021) don't have a DeWitt Clause.

Public cloud vendors

Switching from managed database providers to public cloud vendors the DeWitt Clause usage is interesting. From what I was able to find, Microsoft has the worst terms followed by Google:

• AWS (as of 28-Oct-2021) allows benchmark disclosure via the DeWitt Embrace Clause
  
• Microsoft was harder to find docs on. In this undated agreement if you offer a service on Azure that competes with them, then they are allowed to benchmark and disclose the results of your service. This is far more aggressive than any DeWitt Clause variant. But I didn't find any other mention of benchmarks (to allow or prevent them for Azure).
• Oracle (a PDF dated 22-Jun-2020) has a DeWitt Clause and requires pre-approval for disclosure. They also don't allow performing a benchmark without pre-approval.
  
• Google (as of 18-Oct-2021) has a DeWitt Clause and requires pre-approval for disclosure. They  bans competitors from running benchmarks even when not disclosed. They added the DeWitt Clause on 6-Oct-2020.
  

Public cloud benchmarks

Examples of useful public cloud benchmarks that will be prevented in the future by DeWitt Clauses. I will add to this list as I look for them. For now there was just one from long ago that triggered this post:

• Public cloud object storage perf by Zach Bjornson

More on Microsoft

This is the part in the Microsoft agreement that seems aggressive:

If Customer offers a service competitive to an Online Service, by using the Online Service, Customer agrees to waive any restrictions on competitive use and benchmark testing in the terms governing its competitive service. If Customer does not intend to waive such restrictions in its terms of use, Customer is not allowed to use the Online Service.