1

I am playing around with a MariaDB Galera Multi Master Cluster and I have some performance issues which I do not understand.

I have 2 virtual test machines with Ubuntu 16.04 and MariaDB 10.1 installed. I wrote a simple "benchmark" tool which simulates the creation of a contact in an example application. If I run this benchmark on a single machine, I get around 1050 transactions per seconds on each machine. I did not change anything on the default configuration of mariadb.

If I now cluster those two machines together via galera to a multi-master cluster, I get a significant performance drop to around 99 transactions per second.

The two machines are connected via a Gigabit network. I measured the throughput with iperf and its 950Mbits. The ping times are around 0.35ms. The galera configuration is:

[galera]
# Mandatory settings
wsrep_on=ON
wsrep_provider=/usr/lib/galera/libgalera_smm.so
wsrep_cluster_address="gcomm://192.168.193.220,192.168.193.221"
binlog_format=row
default_storage_engine=InnoDB
innodb_autoinc_lock_mode=2
wsrep_sst_method=rsync
wsrep_node_address=192.168.193.220
wsrep_cluster_name="DB Cluster"

I do not wonder about the absolute transactions per second, but about the relative drop from a standalone instance to the cluster. What am I missing?

1
  • please post the benchmark details. How your SQL statements for "contact in an example application" are? Maybe locking problem? – Josefine Feb 10 '17 at 11:32
1

I don't know about your application/workload but generally, you should try with:

  • wsrep_slave_threads=N (Do not use a value for wsrep_slave_threads that is higher than the average given by the wsrep_cert_deps_distance status variable)
  • innodb_flush_log_at_trx_commit=0 (default is 1)
  • Have you tried redirect all writes to one node and leave the other as a hot-backup? Some workloads get a huge improvement by doing this.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.