TCP Congestion Window Size Reduction
 

Me and my colleagues have been trying to optimize the TCP part of Linux networking kernel, and we have found that Linux does not exactly follow the standards (such as Reno and New Reno) referenced by various Textbooks and RFC's. The situation that troubles us the most right now is described below.

1. The receiver sends three duplicate ACKs to the sender.
2. The sender enters fast retransmission phase

<it's all find upto this point...all described in the standards...now the trouble begins>

3. The receiver sends out duplicate ACK and the sender gets it
4. Just after the duplicate ACK has been received, the sender decreases the congestion window size by one. (this phenomenon directly contradicts the standards, which suggest that congestion window size should increase during the fast retransmission phase.)

If someone knows why the 4th phenomenon occurs, please let me know via kimshinhwa@hotmail.com. If someone could give me a pointer to the solution, that will be great as well. Thank you in advance.