Side_1_360

illustrated in Figure 11 (one maximum segment

size is assumed to be equal to 8 kbyte). The con-

gestion window continues to grow until a time-

out occurs or the receiver’s window is reached.

This part of the algorithm is called slow start.

A parameter called thresholdis also used. This

is initially set to 64 kbytes. When a timeout

occurs, the threshold is reduced to half of the

current congestion window. That is, assuming

that a timeout occurs when the congestion win-

dow is 80 kbyte (as in Figure 11), the next

threshold will be 40 kbyte. Moreover, the con-

gestion window is set to one segment size (equal

to 8 kbyte in Figure 11). So after a timeout, the

slow start approach is again applied until the

threshold is reached.

When the congestion window is greater or equal

to the current threshold, its size is increased by

one segment size (i.e. in a linear manner) as

shown in Figure 11.

The two algorithms described above, slow start

and congestion avoidance, are essential parts of

TCP. Slow start and congestion avoidance are

independent algorithms with different objec-

tives. When congestion occurs TCP slows down

its transfer rate of packets. Then the slow start

algorithm is invoked again. These algorithms

require that two variables are maintained for

each TCP connection, a congestion window,

cwnd, and a slow start threshold size, ssthresh.

Then, these operates as follows (ref. [RFC

2001]):

1. Initialise cwndto one segment and ssthresh
   to 65535 bytes.


2. The TCP output routine never sends more
   than the minimum of cwndand the receiver’s
   announced window.


3. When congestion occurs (timeout or duplicate
   acknowledgements), one-half of the current
   window size is saved in ssthresh. Further-
   more, when a timeout occurs, cwndis set to
   one segment.


4. When new data is acknowledged, cwndis
   increased as: i) in slow start (cwndis less or
   equal to ssthrresh) increase by one segment –
   in effect a doubling of the window for each
   round-trip time; ii) in congestion avoidance
   increase by segsize*segsize/cwnd(segsizeis
   the segment size), which is a linear growth.

5.4 TCP Timers

TCP works with several timers, at least concep-

tually. The retransmission timeris started when

a segment is sent. If an acknowledgement is

received before the timer expires, the timer is

stopped. On the other hand, if the timer expires

before the segment is acknowledged, the seg-

ment is retransmitted and the timer is restarted.

A main question is how to set a proper value of

this timer. Setting it too short will result in many

unnecessary retransmitted packets. Setting it too

long will result in long retransmission delays

and possible low throughput. Therefore, it is

desirable to have an algorithm that dynamically

adjusts the timer value based on measurements

of the round-trip delay. A variable, round-trip

time (RTT)is maintained by TCP, which is an

estimate of the round-trip delay. Whenever an

acknowledgement arrives, the TCP entity mea-

sures how long the acknowledgement took. It

then updates RTT, for the next interval i+ 1,

using a smoothing factor, say a:

RTTi+1= a⋅RTTi+ (1 – a) ⋅T

where Tis the last measured round-trip time.

Typically (ref. [Tane96]), ais 7/8. Then, TCP

commonly uses a value b⋅RTTas the value of

the retransmission timer. One proposal is to set

bin proportion to the deviation of the round-trip

time. However, to simplify the calculations the

deviation, D, can be updated as:

Di+1= c⋅Di+ (1 – c) ⋅|RTT– T|

where cis a smoothing factor. A commonly used

(ref. [Tane96]) value of the timeout is then

RTT+ 4 ⋅D.

When a timeout occurs, one has to decide how to

update the variables. According to the so-called

Figure 11 Illustration of the
congestion algorithm; slow
start, thresholds (maximum
segment size assumed to
be 8 kbyte)

slow start slow start

threshold

threshold

timeout

congestion window (kbyte)

88

80

72

64

56

48

40

32

24

16

8