We've been working with the ABCGs and have found some of their foibles. Specifically, when they rebalance, they only consider the instance count, so the device counts may be heavily skewed. I'm working on a rebalancer that does a better job splitting the load. dataSources that use batchscripts to collect are also VERY heavy handed on the collector they run from. Keeping an eye on # of batchscripts specifically can help show you if your environment is truly balanced. Ours ended up filtering all of the hyper-V hosts to one collector and all of the VMs to the other through rebalanacing... so when a DS fires on one VM, it fires on all of them and since they're all on a single collector, there is no load balancing done. Keep an eye on which resources end up on which collector... you may end up having to increase the balance threshold to prevent it from balancing, then manually move a set of them before lowering the threshold back down to maintain. I go through and force a rebalance about once a week.
Ideally, the ABCG would sort devices by number of instances reporting to each, then tack each next device and hand it to the next collector in the group in a round robin fashion to get true balance of the load. This is the basis of the rebalancer I'm working on. ABCG has been the majority of my time over the past few weeks trying to prevent resource exhaustion on the collectors. We're a very heavy monitoring shop and we're finding the limits of LM.
I have another thread I posted on that has a calculator for getting your threshold to actually balance your instances. I'll have to dig that up.