Galera features beyond just HA
Some of the neat features of Galera are, but are not limited to, multi-master replication, a lightweight implementation of replication and zero failover times due to the multi.master ability. This is not a complete HA solution though, just a component of it, we still need to add some monitoring and failover mechanisms, but as Galera is multi-master this is greatly simplified and can in many cases be handled by the driver or the application with little overhead.
Now, the replication in Galera is synchronous, so that should slow things down a bit, right? Well, yes, but on the other hand Galera can use multiple threads to apply data on the slave, so that should compensate for that somewhat. And how does it compare to MySQL Semi-synchronous replication, which on paper should be that much different?
So I was curious about the multi-threaded apply on the slave that Galera supports? Could this be the multi-thread apply that MySQL has been waiting for all this time? (No, the schema parallel implementation in MySQL 5.6 doesn't count in my mind). So I set out to try this, and this is my thinking:
- The parallel nature of this should be best exposed when you have many small transaction, so each INSERT is a single row, autocommit transaction.
- For the sake of the test, remove as much InnoDB overhead as possible and run on Ramdisk (tmpfs)
- The schema should be simple
- Simple INSERTs are to be tested, nothing else
- Multiple INSERT threads.
- Multi-master operation, but no conflicts.
CREATE TABLE `tab1` ( `c1` int(11) NOT NULL, `c2` char(100) DEFAULT NULL, PRIMARY KEY (`c1`) ) ENGINE=InnoDB DEFAULT CHARSET=latin1And the data to be inserted is
- column c1 - Unique sequential integer.
- column c2 - A random string of characters of 5 to 100 characters length.
- 1.000.000 rows are inserted using 400 threads (200 on each MySQL server).
InnoDB was standard configured here, nothing special, and Galera was using 16 apply threads on the slave, which is probably excessive for this use case. Both MySQL and Galera was using two MySQL servers on the same box.
- MySQL with semi-synchronous replication ached some 4.830 INSERTs per second.
- Galera achieved some 12.987 INSERTs per second, nearly 3 times the performance!
(Note: Opinions expressed in this article and its replies are the opinions of their respective authors and not those of DZone, Inc.)