Date  : Fri, 22 Aug, 2003
Lecture No. 11    (TCP RTT Estimation, Fast Retransmit, Fast Recovery )
Author :  Rupesh R. Mehta

---

 Basic Terminology

• Round Trip Time (RTT) is defined as time between packet send  and  it's Ack is received . 
• Congestion Window (CWND) is sender side limit on amount of data transmission on network before receiving ACK .
• Retransmission Time Out (RTO) is the time set for ACK of particular packet to be received , after this time packet is declared as lost and is retransmitted. 
• Duplicate ACK (DUP ACK) is ACK send by TCP to other end telling that a segment might have been lost and telling it the value of next expected sequence number.
• Slow Start Threshold (ssthresh) is parameter (or value ) used in controlling data transmission.

---

•    In Round Trip Time Estimation,  Smoothed RTT is given as follows :

                             SRTT = alpha * SRTT + (1-alpha) * RTT_i

                In this calculation the value of alpha is choosen to be 7/8 as asssigning  low value gives emphasis on recent RTT, so sudden spikes in RTT value may affect SRTT . Keeping higher value emphasises on old SRTT value and thus sudden spikes will not affect SRTT value very effectively.

•   Initially, the value for RTO is proposed as : RTO = beta * SRTT. But there was problem in choosing beta, Less Value of beta gives    around 50% wrong estimation that packet has been lost and High value of beta makes waiting for long time for lost packets. so,                                                          RTO =  SRTT + 4 * RTT_var      
  

           is taken as value for RTO.     

•     RTO Backoff is technique which keeps on increasing the RTO value exponentially whenever there is retransmission, but limited to few minutes.

•      RTT Sampling Ambiguity is a problem, arises from the way samples are taken for the calculation of the average: TCP is unable to distinguish two separate acknowledgements for the same sequence number. Therefore, we have the following problem:

                    A timeout occurs before an ACK is received, and PKT1 is retransmitted. The ACK for the first PKT1 arrives a bit later and the source measures a wrong value for the RTT.


                                              fig.  RTT is considered as difference between most recently retransmitted and ACK received for the Original packet.    

    Thus  using the original transmission for RTT measurement, causes Over-Estimation, whereas using  the most recent retransmission for RTT measurement causes Under-Estimation and thus gives very negative effect. ( As less RTT tends to reduce in RTO which further causes estimating packet lost even if packet is not lost and thus causes retransmission.) 
  Ignoring RTT sample also leads to the problem of continuos retransmission (reason given below), as RTT changes due to several reasons such as congestion in network , change in route depending on some policy.

• Karn's Algorithm :  

                            Rule 1: Ignore measured RTT for retransmitted packets. 
                                         This removes ambiguity from RTT measurements.

                            Rule 2: RTO should be doubled after retransmission.
                                        This is called "Exponential Back-off".

Why is Rule 2 necessary?
When RTO is smaller than Real RTT..
        If only Rule 1 is applied, TCP will use previous RTO forever. Thus all packets will be retransmitted.
        No RTT mesurement happens.

•       Using Timestamp to resolve ambiguity is also a good solution. As the other end will copy the value from the segment being acknowledged and put it in the ACK and send back to sender , then it is easy to calculate RTO. But this scheme involves overhead (i.e. can not compress header).

  FAST RETRANSMIT / FAST RECOVERY  :

• Fast Transmission :                                                                                                                                                                                                            When TCP detects segment loss using retransmission timer, then ssthresh is set to half of CWND. i.e.  ( ssthresh = CWND / 2  )  and CWND is set to 1-full size  segment. Now instead of waiting for retransmission timer to get off , TCP detects packet loss by looking for packet re-ordering and retransmit the lost packet. This scheme is called as " Fast Retransmit". In this algorithm , TCP receiver sends an immediate duplicate ACK on out-of-order segment arrival.  The other end TCP deduce from small number ( normally 3 ) of consecutive duplicate ACKs that the segment has been lost and deduces the starting sequence number of missing segment. The missing segment is retransmitted.
  

•  Fast Recovery :                                                                                                                                                                                                                   This algorithm says that after Fast Retransmission Algo. (i.e. after the missing segment has been retransmitted), Congestion Avoidance but not the slow start is performed. This is an improvement that allows higher throughput under moderate congestion, especially for large window. This algorithm governs the transmission of new data until a fresh (new) ACK arrives.
  

              
The Algorithm works as follows :  

              If  Number of DUP-ACK received  = 3
                           ssthresh =   CWND / 2;
                            Retransmit the lost segment ;
                            CWND = sshthresh + 3 ;       /*  3 is for 3 DUP-ACKs */

            For each next DUP-ACK
                          Increment CWND  by 1.
                                  /* i.e. CWND = CWND + 1. This inflates the CWND in order to reflect segment that has left the network */

            If   allowed by CWND and Receiver's Advertised Window    
                         Transmit a segment (if any....)                                              

            If  Not DUP-ACK                        /* i.e. New/Fresh  ACK  */
                          CWND = sshthresh         /* Deflating the window */

           Call Congestion Avoidance Algorithm           


 Congestion Avoidance (CA) :

                  Consider,  sender's  CWND = W . Packets going to the receiver have sequence numbers ( U , U+W) , now consider that the packet having sequence no. 'U'  gets dropped, [ U, U+W)  packets are still in transit. On packet loss detection, sender will reduces it's CWND size to half of its original (i.e. W/2).    The receiver will send  W-1 DUP-ACKs in single RTT. Along with retransmitting lost packet, sender will send  at the max. W/2 -1  new packets (as CWND = W/2 ). Thus total number of packets sent  are [ U, U+W/2+W-1) . If only 'U' packet had  lost then the new ACK will ask for the U+W .