Checkpoint sorting and kernel interaction
When PostgreSQL writes changes to disk during a checkpoint, it now does so in a more orderly way to ensure that writes are more sequential than earlier. This is done by sorting blocks before sending them too. Random writes will be dramatically reduced this way, which in turn leads to higher throughput on most hardware.
Sorted checkpoints are not the only scalability thing to make it into 9.6. There are also new kernel write-back configuration options: what does this mean? In case of large caches, it could take quite a long time to write all changes out. This used to be especially nasty on systems with hundreds of gigabytes of memory because fairly intense I/O storms could happen. Of course, the operating system, level behavior of Linux could be changed using the /proc/sys/vm/dirty_background_ratio command. However, only a handful of consultants and system administrators actually knew how to do that and why. The checkpoint_flush_after, bgwriter_flush_after, and backend_flush_after functions can be used now to control the flush behavior. In general, the rule is to flush earlier. Still, as the feature is new, people are still gathering experience on how to use those settings in the most efficient way possible.