Active Queue Management

Question

Was reading through one of the papers on Active Queue Management.Just hit upon this paragraph, which I couldn't understand.Got confused at this part:

https://queue.acm.org/detail.cfm?id=2209336

Figure 4 shows the queue vs. time for a TCP receiver that sends one ack per window rather than one per packet. This is a legal, though brittle, operating mode used by some commercial operating systems with very high per-packet overhead to reduce the number of acks generated so as to perform credibly in benchmark tests. This kind of queue length variation is also seen in Web traffic—a superposition of mostly small transfers that are essentially all startup transient. Since this ack policy means a full window of packets is always delivered as a back-to-back burst, the initial turn-on transient repeats every RTT and the buffer is always occupied. The window exactly fills the pipe, however, and there is no excess queue, so any attempt to reduce this queue will result in poor utilization of the bottleneck. Thus, occupancy time contains no information about excess queue.

1). My understanding is Ack are send per packet always. if not what does it mean by sending ack per window as well? 2). what does this statement mean "Since this ack policy means a full window of packets is always delivered as a back-to-back burst, the initial turn-on transient repeats every RTT and the buffer is always occupied"?

Answer 1

Q1) A host should send ACKs as soon as possible = on a per-packet basis. However, as the text mentions it's not in violation against RFCs to only send one ACK at the end of the window.

Q2) The sender sounds out the whole window and only when the window is finished the ACKs arrive and a new window is scheduled. This is in contrast to the normal, sliding mode of the send window. This burst is usually only seen at the very start of a transmission where TCP parameters have to align themselves. The 'ACK only at end of window' prevents this alignment.

Pretty obviously, this ACK scheme destroys the purpose of the sliding window scheme. Essentially, it limits the throughput to one window per RTT. It's quite like a receive/send window of 1 with a packet the size of the window.