MariaDB 5.5 performance on Windows

We haven’t posted any Windows benchmarks for a while, and MariaDB for Windows contains some specific improvements which might not be widely know since we haven’t talked much about them yet. This post is an attempt to fix that. We’ll also share current MySQL 5.5 numbers.

My setup is an 8 core 2 socket server (yes, a little bit dated for today, but it is the best machine I have at my disposal), 10K SAS disks with RAID1. I ran sysbench 0.4 single table / 1,000,000 records. I ran the benchmark over a network, with the number of concurrent clients ranging from 4 to 4096.

Here is what OLTP-readonly throughput looks like:

  • For most of the tests, MariaDB’s throughput is approx 10% higher than MySQL’s
  • For 4096 concurrent clients, MariaDB’s throughput is better than MySQL 5.5 by 476% (2382 vs 413 TPS).

Many people rightfully remark that throughput does not necessarily represent performance as a whole, and that in OLTP benchmarks, fairness also matters. It actually often matters more than throughput. I agree, so below are the results for 95% of response time, meaning that 95% transactions were finished under the given time.

OLTP Readonly response time, 95th percentile

Concurrent clients 4 8 16 32 64 128 256 512 1024 2048 4096
MariaDB 4.87 6.81 8.83 12.35 22.12 43.56 90.35 180.57 619.05 1003.88 1965.77
MySQL 4.86 7.14 9.96 16.21 37.39 101.33 238.89 499.63 971.07 2241.83 25215.29

As the above table shows, MariaDB 5.5 outperforms MySQL 5.5 in both row throughput, and in fairness. Fairness (response time) results are in fact more impressive than the row throughput.

But, why does MariaDB perform better than MySQL 5.5 on Windows in readonly tests? Why it is at all better than 5.5? It is, after all, based on the same codebase, so performance should be about equal. The answer to this is not our optimizer enhancements, nor XtraDB serving as InnoDB in MariaDB. The answer is the MariaDB threadpool. This is my feature in 5.5 and I’m glad it performs so well, and I’m glad that it is switched “on” by default on Windows (so, Windows users, you do not have to do anything to switch it on).

Still, why does using threadpool result in such better performance? The answer is that the implementation delegates all responsibility for sizing the pool and running callbacks to the Windows native threadpool, so it is a sort of black box inside the OS that just delivers good results. The heart of the Windows native threadpool, if you want to know, is the good old IO completion port, something that Windows has had since the NT 3.5 days. It’s a unique Windows feature, and something that any server application running on this platform is using or should be using. And the tricks that make it run well are:

  • It does not let too many threads run “on CPU” at the same time (this reduces context switches). Reducing context switches is, in my experience, the single most important factor to increasing throughput.
  • It activates threads waiting on completion LIFO order — hot threads remain hot, which reduces cache misses
  • IO completions are processed in FIFO order — this is why fairness performance is so good
  • And yes, it naturally reduces contention on hot locks.

Enough about readonly performance already, and about threadpool.

The next interesting question is whether MariaDB can do as well on write workloads. So let’s run the sysbench in the ultimate “write-only” mode, also known as update_non_index (every query only increments a non-indexed int column). To maximize the write throughput, I set the innodb_flush_log_at_trx_commit parameter to 0, so the log is flushed once a second rather than each time at transaction commit.

Here are the results:

OLTP write-only (update_non_index/flush_log=0) throughput:

This looks pretty good. The difference could be a combination of many factors. XtraDB write performance, group commit, and threadpool could all have a positive (for MariaDB) influence on the results. However I think the Windows version of MariaDB’s asynchronous IO optimization (it, again, uses completion ports) is the main reason for the big difference here. I, actually, first saw the difference when I was implementing the feature.

All IO related parameters and Innodb parameters in the test above are default. The result looks almost too good to be true, so I really wanted to bring MySQL to the point it can achieve better IO throughput by varying innodb_io_capacity and/or innodb_write_io_threads, to no avail. Does someone know the right trick here?

OLTP writeonly (update_non_index/flush_log=0) response time, 95 percentile

Concurrent clients 4 8 16 32 64 128 256 512 1024 2048
MariaDB 0.33 0.63 0.75 1.06 1.94 3.85 8.25 21.38 129.79 274.40
MySQL 0.32 0.61 0.73 1.61 7.62 26.82 96.45 219.29 661.19 2723.36

For completeness, here are database parameters I’m using.

[mysqld]
sql-mode="NO_ENGINE_SUBSTITUTION"
back_log=500
user=root
port=3306
max_connections=4096
max_connect_errors=5000
max_prepared_stmt_count=50000
table_cache=2048
transaction_isolation=REPEATABLE-READ
loose-skip-external-locking
innodb_status_file=0
innodb_data_file_path=ibdata1:200M:autoextend
innodb_buffer_pool_size=1G
innodb_additional_mem_pool_size=20M
innodb_log_file_size=650M
innodb_log_buffer_size=100M
innodb_support_xa=0
innodb_doublewrite=0
innodb_flush_log_at_trx_commit=0
query-cache-size=0
query-cache-type=0
symbolic-links=0
skip-grant-tables