+ * Most of the I/O is now buffered, which results in a pretty
+ large speedup when running under MS Windows. (Craig Barratt)
+
+ * Optimizations to the name-handling/comparing code have made
+ some significant reductions in user-CPU time for large file
+ sets.
+
+ * Some variable-type cleanup that makes the code more consistent.
+
+ * Reduced memory requirements of hard link preservation.
+ (J.W. Schultz)
+
+ * Implemented a new algorithm for hard-link handling that speeds
+ up the code significantly. (J.W. Schultz and Wayne Davison)
+
+ * The --hard-link option now uses the first existing file in the
+ group of linked files as the basis for the transfer. This
+ prevents the sub-optimal transfer of a file's data when a new
+ hardlink is added on the sending side and it sorts alphabetically
+ earlier in the list than the files that are already present on the
+ receiving side.
+
+ * Dropped support for protocol versions less than 20
+ (2.3.0 released 15 Mar 1999) and activated warnings for
+ protocols less than 25 (2.5.0 released 23 Aug 2001)
+ (Wayne Davison and J.W. Schultz, severally)
+
+ * More optimal data transmission for --hard-links (protocol 28).
+
+ * More optimal data transmission for --checksum (protocol 28).
+
+ * Less memory is used when --checksum is specified.
+
+ * Less memory is used in the file list (a per-file savings).
+
+ * The generator is now better about not modifying the file list
+ during the transfer in order to avoid a copy-on-write memory
+ bifurcation (on systems where fork() uses shared memory).
+ Previously, rsync's shared memory would slowly become unshared,
+ resulting in real memory usage nearly doubling on the receiving
+ side by the end of the transfer. Now, as long as permissions
+ are being preserved, the shared memory should remain that way
+ for the entire transfer.
+
+ * Changed hardlink info and file_struct + strings to use
+ allocation pools. This reduces memory use for large
+ filesets and permits freeing memory to the OS. (J.W. Schultz)
+
+ * The 2 pipes used between the receiver and generator processes
+ (which are forked on the same machine) were reduced to 1 pipe
+ and the protocol improved so that (1) it is now impossible to
+ have the "redo" pipe fill up and hang rsync, and (2) trailing
+ messages from the receiver don't get lost on their way through
+ the generator over to the sender (the latter mainly affected
+ hard-link messages and verbose --stats output).