I suspect this is just bursting/clamping, as you suspect, but you
may also want to investigate traffic shaping at your end. I've found
I get much better *receive* throughput if I limit my *transmit* rate
to less than nominal maximum. Presumably, this has to do with the
fact that the feed is asymmetric; I can receive much faster than I can
send, and so the send channel becomes congested and that impacts TCP
ACK or other protocol control messages.