Note from the author of Backshift, Dan Stromberg, August 2015

I'm the author of backshift; I just noticed your useful backup performance comparison.

Backshift is never fast - it's more about being frugal with diskspace. You're correct that this is in part because of network filesystem I/O, but a lot of it is because of the block-chopping algorithm and (slow but thorough) xz compression.

However, on your initial fullsave, backshift is quite a lot faster if you run it on Pypy, because the first fullsave is very CPU-intensive.

Subsequent backups tend to be faster on CPython with Cython, because subsequent backups tend to be more I/O intensive, and CPython still beats Pypy for I/O performance.

Examples of configuring backshift for use with Pypy or CPython+Cython are at http://stromberg.dnsalias.org/~strombrg/backshift/documentation/for-all/installation.html

Graham's note, February 2016:
There is further detail from Dan at the following link.
http://strombrg.blogspot.com.au/2015/10/backshift-is-deduplicating-backup.html