The update to the latest version of NR (7.1.51) should resolve that give you the v6 behavior back. In V7 they changed how jobs were handled and if a slave was already working on a chunk, no other slave would touch it, even if the other slave was much faster or had more cores in its pool. This was done so another job in the queue would be able to start processing on the free slave. Nice theory, but it only works efficiently if all your slaves are about equal in power, which for us is definitely not the case. We have two machines with 48 cores each, but then two 8 core machines, so with 7.1.41 every job ended up sitting there waiting on the 8 core machines while the other 96 cores sit idle if there wasn't another job right behind it. This is also why some of us have requested that the 'dicing' size (size of the chunks) is controllable, since we have some large sperhical VR's where a single chunk took 2 days to process on the slower machines.
7.1.51 basically rolled that back to the v6 behavior where it should resend the chunk to the (faster) idle slave and race to see who can finish first before it moves to the next job. Keep in mind that with 7.1.51 you will have to update every slave and every user machine that uses Network Rendering. v7.1.51 is not backwards compatible with 7.1.41.