Sysbench complex test
Ok, preparing the DB resulted in mixed results, but what about running the sysbench complex test against it?
At first, results are very low for all the contenders. However, you should remember that this benchmark runs concurrently to the others (fileserver, varmail and webserver), so the host had a very hard time keeping pace with the outstanding I/O operations.
Basically, any setup apart BTRFS and ZFS give similar performance. Even ZFS can be deemed acceptable. However, BTRFS is a complete fiasco here: it is 10 times slower than ZFS.
Now, have a look at latency:
A very similar pattern emerge. In other words, don't use BTRFS for a PostgreSQL installation, nor for any storage subsystem backing a PostgreSQL VM.