Dicts are being used more and more in our I/O path (e.g. for xdata) so performance is a significant concern and our current implementation allocates data_pair structures separately from the dict itself. Since many dicts are used to hold a very small number of items - often just one - we can optimize by allocating the first data_pair as part of the dict structure.
http://review.gluster.com/3040 is commited to master branch