 |
Date Posted:
Thursday June 26, 2003 08:54:55 AM
|
|
I am Chong Kim (the author of this article).
I have a proposal to make regarding what to do when ASC approaches FZB. We know that TCP aborts the connection when the data octet(s) has been retransmitted a number of times, i.e. duplicates, but its acknowledgement does not arrive. Let's say that the sending TCP may wait for the maximum time T since it emitted the data octet(s) for the first time before the abortion of the connection. During this time T, a number of retransmissions take place if an acknowledgement does not arrive. What we could do is that, when ASC is at least T units in t away horizontally from FZB, the sending TCP initiates the resynchronization of the sequence number with the receiving TCP. This ensures that ASC never enters FZ. That is, either the sender aborts the connection because none of the resynchronization packet of the sender and its possible retransmissions has been acknowledged before time T elapses or the resynchronization is successful before ASC meets FZB. With this scheme, we can do some arithmetic to raise the possible minimum sequence number of the data octet in (1) in "The evaluation of m - case (1)", which will lead to a smaller lower boundary of our m (i.e. Let S = the maximum waiting time after the last retransmission of a packet. (i) If we allow the sender to initiate the resynchronization before all the data octets emitted earlier are acknowledged, the sender emits the resynchronization packet at least T away horizontally from FZB. For this, if MSSW >= a(T - S) + 1, add (a×T - 1) to (1) in "The evaluation of m - case (1)", which will make the lower boundary of our m smaller by (a×T - 1 ) . In this case, do not forget to replace MSSW with (a×WT+1) or a(T - S )+1. If MSSW < a(T - S) + 1, add (MSSW + a×S - 2) to (1) in "The evaluation of m - case (1)", which will make the lower boundary of our m smaller by (MSSW + a×S - 2). Of course, with the same scheme, there are other ways to make the lower boundary of our m smaller. For example, the sender may emit the resynchronization packet at least (T - S) away horizontally from FZB. In this case, ASC may enter FZ but, while ASC is progressing horizontally in FZ, no data octet whether duplicate or not is emitted. For this, if MSSW >= a(T-S) + 1, add a(T-S) -1 to (1) in "The evaluation of m - case(1)", which will make the lower boundary of our m smaller by a(T-S) -1. If MSSW < a(T-S) + 1, add (MSSW-2) to (1) in "The evaluation of m - case(1)", which will make the lower boundary of our m smaller by (MSSW-2). (ii) Suppose that all data octets should be acknowledged before the sender initiates the resynchronization. Then when ASC is at least 2×T away horizontally from FZB, the sender sends no more new data octets and waits until all the data octets are acknowledged, after which the sender emits the resynchronization packet while ASC is at least T away horizontally from FZB. ) I have assumed that the horizontal width of FZ is relatively large (with T relatively very small to the t range ) so that the sender can not afford to remain idle while ASC progresses horizontally in FZ until it exits from FZ (i.e. I have heard that, in one implementation of TCP, the horizontal width of FZ is set to a value larger than 3 minutes! Suppose that the horizontal width of FZ is 4 minutes and the clock cycles every 10 hours. Suppose that ASC can enter FZ horizontally. Then, with a security system using TCPs that run day and night, the system may stop working for 4 minutes every some hours which is smaller than a small multiple of 9 hours and 27 minutes. What about a bank service till machine that occasionally stops working for 4 minutes every few days or once a day? ). Or if the maximum waiting time S after the last retransmission of a packet is larger than the horizontal width of FZ, we do not need FZB at all. In this case, the sender initiates the resynchronization when ASC is horizontally T away from ISC (after the clock cycles). FZ is implicit in this case. I do not suppose that such S is implemented. If such S is implemented, we can also make the lower boundary of our m smaller using the previous arguments above.
Also in "The evaluation of m - case (1)", we do not consider any data octet, whether duplicate or not, emitted before t = t1 - MSL - C - A + B exclusive because such a data octet will not exist in the network after t = t1+B inclusive ( i.e. after the new connection opened at the receiver ).
|
