This is the story of a server that was getting a bit short in memory.
Change request issued to the hosting company, please throw four more GigaBytes of RAM in that guy.
One of the processes indexes the content of a database table (hosted on /dev/sda) and store it on /dev/sdb on an hourly basis, in order to speed up some searches.
Impact of the increased RAM over this process?
You guessed it, the RAM upgrade happened on the 16th.
What about the overall memory utilization?
The dreadful committed memory got reduced to much more sane levels, whereby the server no longer takes the risk of falling over if processes start actually using the memory they (m)allocated.
I venture that the increased violet-ish part of the graph explains in part the improved I/O behavior.
OS caching is great.