Performance of the parallel slave critically depends on how different computation-wise
individual transaction of the same group commit are.
Note that the slave starts executing them in parallel in whole unlike the master who merely
commits such way transactions prepared at about the same time. The best slave performance therefore can be achieved when
the master binlog group transactions are of equal sizes. This assumption is unpractical, naturally.

There are some ideas, partly discussed on M18, how to overcome this obstacle.

See for more HLD.txt and the following discussion in https://lists.launchpad.net/maria-developers/msg11524.html.