David S. Miller wrote:
On Tue, 25 Nov 2003 08:47:29 -0500 Ted Duffy <tedward@egenera.com> wrote:
However, I think this disparity would cause cwnd to be increased too aggresively given the true knowledge the TCP is gleeming from the network based on successful acks. In this case, the ack of a 50 byte message results in cwnd being increased a full packet size (1500 bytes, 9k, or whatever else the MTU might be).
Your statements would be very interesting if routers dropped bytes
instead of packets, however they don't.
Given conventional, relatively simple networks, this is fair enough. Although given the argument that routers drop packets and not bytes, one has to wonder why TCP is implemented at all with octet counts. Keeping track of the congestion window on an octet basis certainly doesn't hurt anything and is more in-line with the rest of the Linux TCP implementation as well as the RFCs.
I'm not too familiar with a wide variety of networks, their limitations, and practical uses, but it seems to me that it is not too far fetched for a sufficiently complex network to encounter congestion with large packets, while having no problem with smaller packets. Increasing the congestion window based on packet count when little packets are acked can then result in inaccurate congestion knowledge.
Note, however, that the issue of managing the congestion window on a per-packet basis is currently all theortical, from my point of view. While I have read comments about this in assorted venues, my personal experiences, so far, have not found any issues related this.
Ted
- : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
