Only the last is around pg, but I think they all may still be applicable.
The kernel docs in Documentation/vm/transhuge.txt have an explanation of the metrics
We hadn't been having the issue that much until a few weeks ago, when we started using the rest of our free memory for page cache.. my thoughts were if we have no more memory that's totally free, it might be doing compaction more. That lead me to find how often compaction is happening, but like I said I don't know how to tell how *long* it's happening - someone who knows systemtap better than I might be able to help with the collection of that info assuming the right systemtap events are there.
