For last
years the protocol of control of transmission (TCP) in a combination to the
Internet by the protocol (IP) of steel by the most widespread form of organization
of networks between computers. In connection with conducted development of
high-speed networks and applications which use TCP, the problems have behaviour
ÒÑÐ the increased importance. To behaviour (i.e. productivity, amount of the
lost packets, the stability etc.) during an initial stage is necessary to
give separate attention, as the majority of the applications use ÒÑÐ for short
transmissions, which are finished before as ÒÑÐ will pass to steady behaviour.
One of the approaches in research of problems of behaviour consists in study
of job of concrete realization TCP. Thus in the given job, with the help of
simulation, the behaviour of several realizations ÒÑÐ (Tahoe, Reno, SACK,
Vegas) is in details studied. On the basis of these supervision some simple
changes are offered which at the certain situation essentially improve behaviour
of the protocol.
The modern
realizations TCP contain set of algorithms aimed at control by a network overload
at maintenance of good productivity of the user. The early realizations TCP
followed model using cumulative positive acknowledgement and the requirement
of the expiration of the timer for a repeated sending of the data, lost at
transportation. These TCP made for minimization of an overload in a network
a little.
The realization Tahoe TCP has added set of new algorithms and improvements
to earlier realizations. The new algorithms include Slow start, Prevention
of an overload, and Fast repeated transmission. The improvements include updating
in an estimation of time of circular reversal for installation of meaning
of a waiting time of repeated transmission. All updatings were described in
[Jac88].
To algorithm of Fast repeated transmission the soul interest in this job is
shown, because he is modified in the subsequent versions TCP. With Fast repeated
transmission, after reception of small number of double acknowledgement for
one packet TCP (dup ACK), the source of the data concludes, that the packet
was lost and repeatedly passes a packet expecting the expiration of the timer
of repeated transmission, that conducts to higher use of the channel and productivity
of connection.
Reno TCP saves the expansions
which have been switched on in Tahoe, but changes operation Fast Retransmit
adding Fast Restoration [Jac90]. The new algorithm prevents a way of communication(connection)
("channel") from a finding in an empty estate after Fast repeated
transmission, thus leaves from requirement of Slow start for filling the channel
after single loss of a packet. The fast Restoration assumes, that everyone
received dup ACK represents one packet left channel. Thus, during Fast Restoration
the sender TCP is capable to do intellectual estimations about quantity of
the sent data.
The entrance in Fast Restoration by the sender TCP occurs after reception
of an initial threshold dup ACK. This threshold normally known as tcprexmtthresh,
is normally established equal to three. As soon as the threshold dup ACK is
received, the sender repeatedly passes one packet and reduces its window of
an overload by half. Instead of Slow start, as in Tahoe TCP, the sender Reno
uses additional ward dup ACK to synchronize the subsequent leaving packets.
The algorithm Reno of Fast restoration is optimum only at individual losses
of packets. The sender Reno passes no more than one packet for one time of
circular reversal. Reno considerably improves behaviour in comparison with
Tahoe TCP when there is a loss of one packet, but can essentially worsen a
situation if there was a loss of several packets within the limits of one
window. The given problem is illustrated at simulation.
The algorithm TCP SACK
uses a field of "Option" of heading of the frame ÒÑÐ for the additional
information on the received packets by the station - addressee. If there was
a loss, each segment dup ACK sent by station by the addressee contains the
information on the frame called a sending of the given segment. Thus sender,
having received the given frame has the information not only about that what
frame was lost, but also and about that what frames successfully have achieved
the addressee. Due to this the unnecessary repeated sending of segments is
avoided, is successful bufferised on the side of the addressee.
As well as Reno, TCP SACK enters into a mode of Fast restoration at reception
3 duplicated acknowledgement (dup ACK). During Fast restoration the sender
supports variable pipe, displaying number of packets which are taking place
in a network. Given variable each time is increased, when the new segment
was sent and decreases, if the next acknowledgement was received. The transmission
of a new packet to a network is solved if meaning pipe less window of an overload.
The sender also supports a data structure scoreboard, which remembers acknowledgement
from an option SACK of arriving acknowledgement. If to the sender the transmission
is solved, he passes the following packet from the list of the packets considered
lost. If such packets, the new packet is sent.
On the basis of features
of realization of algorithms submitted above the model for each protocol is
under construction. An important point in construction of model is the superexact
display of behaviour of a puncture in a real network so that the results received
by an experimental way could be transferred to a real network.
In a fig. 1 the simulated topology in a network used in the given job is submitted.
Fig. 1. Topology of model
In a fig.. 1 the circle represents a sluice with the final buffer such as FIFO, and the squares represent host of the sender and addressee. The used time of circular reversal is established equal 1 msec. For singleness all models submitted in the given job use the addressee, which sends a broadcast television complex on each received segment. A bottleneck in a network is the sluice, overflow of which buffer and represents an overload in a network.
In a course of simulation
the following results were received.
· at exponential increase of a window during Slow start in the protocol ÒÑÐ
Tahoe, if the size of the buffer is less than 1/3 bandwidth-delay product,
there is an overflow of the buffer, that results in the second Slow start
and reduction of productivity.
· TCP Reno tries to avoid this effect reducing a window twice at detection
of loss. Though it also provides the best productivity at perfect conditions,
TPC Reno in the present form is too vulnerable what to be replacement Tahoe.
· Use of the "rough" timer results in catastrophic loss of productivity,
if timeout is the identifier of loss of a packet. Use of more complex processing
of an estimation of time of circular reversal therefore is necessary.
· The ideal behaviour at multiple loss of packets represents TCP SACK, however
use of the given realization is possible only in those cases when both ends
of connection (both sender and addressee) support the given option.
· At absence of an option SACK the situation of repeated fast transmission
is possible, that results in significant reduction of productivity. In order
to prevent the given effect it is necessary in addition to use variable "send_high",
reflecting the maximal number of a sent packet, or, that is more universal,
to analyze a departure time of the false lost packet. Last variant is more
combined as is connected to an exact estimation of time of circular reversal.