Interesting information, thanks for sharing.
Although not directly related, this reminds me of a discussion I had with some Microsoft folks quite a few years back now, around the transition from Server 2003 to Server 2008. The issue was a fairly significant performance degradation when a UNC path pointed to a file share on a local machine compared to a UNC path on a remote machine and especially when compared to I/O on the local file system(s). It seemed the Windows kernel or network redirectors were not designed, at that time, to bypass or short-circuit the network stack when a network path pointed to a local file system. This was in contrast to many Unix or Linux systems that would/could optimize that path through the system stack. This impacted our ArcGIS for Server deployments because the common practice in my organization back then was UNC paths for all data, local or otherwise. It would be fun to revisit that topic and do some benchmarking today, hopefully that problem is in the past, hopefully.
Of course local disk I/O is faster than network I/O, especially with SSDs coming on strong, but gigabit Ethernet is respectable. I guess the 50% performance penalty surprises me, a bit. Have you tested on 10GbE networks?
