An abridged version available under the same title in Proc. ICNP 2003, pp.95-106, Nov 2003.

The Tech Report that proposes DSACK-TAES to make TCP perform well under reordering networks, and support the proposal with simulation.

DSACK can report dup packets to the sender to help sender learn about a spurious retransmission. Such spurious retransmission is called false fast retransmit in the paper, as it is caused by the fast retransmission algorithm in TCP. The paper proposed that, when FFR is detected, the window reduction can be undone as there was no congestion.

To prevent future FFR, one can set a larger dupthresh, so that FR would not kick in before enough dup ACKs to confirm a lost. But a large dupthresh makes the TCP sender react slower after real packet drop. In FR, RFC3042 allows a limited transmit which the sender can send new data packet (under self-clock) once upon several duplicated ACKs received. Default is only for each of the first two dup ACKs. But upon a real loss, this limited transmit can delay the reduction of cwnd.

The core idea in this paper is the DSACK scoreboard. The sender record each FR’s starting time and the amount of cwnd reduction in the scoreboard for the retransmitted packet. If the FR is later found a FFR, the time between the start and the end of the FFR are recorded. Simply speaking, this measures the reordering length, and it aggregates the reordering length samples into a histogram. Then the dupthresh is set using the data in the scoreboard histogram. The histogram keeps the data for the recent confirmed FFR incidents. Base on the histogram, we can deduce the reordering length that fits into a certain percentile of reordring events. Then dupthresh is set accordingly. For example, we found that 90% of reordering involves 8 or less packets, then setting dupthresh to 9 can prevent 90% of FFR. The percentile is called the FA (FFR Avoidance) ratio and this scheme is called DSACK-FA in the paper.

DSACK-FA can be furthered to make the FA ratio adaptive to maximize throughput. The RTO event in TCP is doing more harm. If we created a RTO because of the dupthresh too high, the cost would be losing the opportunity to send a lot of packets. Also, FFR has the similar cost because it sends slower during FR because we can confirm it is an FFR. And the limited transmit has the cost of making the sender idle at the time that it can send packets. So taking these costs, we can adapt the FA ratio such that: (1) Upon FFR, increase the FA ratio by S, (2) upon RTO, decrease the FA ratio by $\dfrac{C_{TO}}{C_{FFR}}S$ where $C_{TO}$ is the cost of RTO and $C_{FFR}$ is the cost of FFR, (3) upon limited transmit-causing idle period, decrease the FA ratio by $\dfrac{C_{LT}}{C_{FFR}}S$. The value of S given in the paper is 0.01. This is called DSACK-TA for Timeout Avoidance.

Moreover, in DSACK, as we have the timestamp reported by the sender, we can differentiate an ACK of retransmitted packet from an ACK of original packet. While Karn’s Algorithm says not to sample RTT during retransmit because of ambiguity, such ambiguity is avoided by the timestamp. Thus we can get a more accurate RTT estimate if we make use of the reported timestamp. Indeed, the Karn’s Algorithm is biased against short RTT paths and gives an over-estimated RTT value. Such improvement is called DSACK-ES, for Enhanced RTT Sampling in the paper.

The simulation in this paper shows that, DSACK-TAES can significantly improve TCP’s througput performance in face of reordering.

Bibliographic data

@techreport{
   title = "RR-TCP: A Reordering-Robust TCP with DSACK",
   author = "Ming Zhang and Brad Karp and Sally Floyd and Larry Peterson",
   howpublished = "ICSI Tech Report",
   institution = "International Computer Science Institute",
   number = "TR-02-006",
   month = "July",
   year = "2002",
}