...
So what I suspect you're actually measuring there is the latency of the mining process... which is consistent with what we've expirenced with P2Pool (5-20 second latencies from ASIC miners are common).
Interesting!
And this is why I like the empirical "block box" approach. I don't care initially what the
mechanism is. I try to find a simple model that explains the effect, and then, later, ask what that mechanism might be.
But now why would the "latency of the mining process" depend on the size of the previous block? That doesn't make sense to me, but we just
showed empirically that F2Pool (but not AntPool) is indeed more likely to produce an empty block when the previous block was large (suggesting that processing large blocks [including the mining process latency] takes longer for some reason).