Once I've finished the NAT module for Cisco's Skinny protocol I'd like to write a "filesystem". My researchers often use raw partitions for saving data -- they transfer a few big datasets and transfer speed wins over the convenience of a general-purpose filesystem. The huge number of virtual partitions allowed by LVM has been a win.
However, filesystems are convenient: Naming, times, access control. What is needed is a filesystem no more complex than dd. Too stupid for general use: bad things happen as the filesystem fragments into holes from file deletion. Interesting for high-performance data transfer and for measuring the overhead of a filesystem compared to dd (which I've read may be as high as 30%, but we will actually be able to know, rather than to suspect).