> But, assume... ACK packets arrive at TCP sender with same rate of DATA
> packets.
> It is easy to guess increasing queue delay.
>
> 1 2 4 8 16 32 64 128 200 200 ....
>
> 1~200 : slow-start
> 200 ~ .. : congestion avoidance
>
> I'm sure, packets are spread out over the RTT during congestion avoidance
> period.
> (--> Because, One ACK packet inject One DATA packet into queue)
>
> How about slow-start ?
> One ACK packet injects Two DATA packets into queue.
> Certainly, only one data packet dequeue and other data packet remains queue.
> Each successive ACKs inject Two data packets.
> If output rate of data packet is equal to receive rate of ack packet, as
> CWND increase,
> queue delay increase, during slow-start mode.
Yes, but in a WAN the bottleneck link is normally not the one that connects
your machine. If your machine is connected to a 10Mb/s Ethernet which is
further connected to a T1 link (200KB/s), then the queueing would occur at
the router that connects Ethernet to T1.
Even if the link connecting your machine is the bottleneck, it would mean
a queuing of 50*1500 i.e. about 75K. The vast majority of TCP receivers in
the Internet advertise flow control windows that are much less than this.
So you're not in danger of buffer overflow due to one TCP connection. However,
several TCP connections can indeed cause overflows.
- Mohit
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:42 EST