IND426 Tutorial 4 (Solutions)

(b) S = 500x1024/(25x106 - 2x106) = 22.3 msec. There is a 25 MB/s output burst of 22.3
msec after the onset of the incoming burst of data.

(Roberto Togneri 1998)

1.

(a)

(b)

(c)

Consider a window flow controlled virtual circuit going over a satellite link. All
packets have a transmission time of 5 msec. The round-trip processing and
propagation delay is 0.5 sec. What is the maximum speed transmission rate in
packets/sec? Find a lower bound on the window size for the virtual circuit to be
able to achieve maximum speed transmission when there is no other traffic on the
link. Transmission time for the acknowledgment is negligible.
Suppose the virtual circuit goes through a terrestrial link in addition to the satellite
link. The transmission time on the terrestrial link is 20 msec, and the processing
and propagation delay is negligible. What is the maximum speed transmission rate
in packets/sec? Find a lower bound on the end-to-end window size for the virtual
circuit to be able to achieve maximum speed transmission when there is no other
traffic on the link. Transmission time for the acknowledgment is negligible and
routers use a store-and-forward scheme.
Suppose that node-by-node windows are used in the two-link system. Find lower
bounds on the window size required along each link in order to achieve maximum
speed transmission, assuming no other traffic on the links. Transmission time for
the acknowledgment is negligible and routers use a store-and-forward scheme.

Solution
(a) Max. Transmission speed = 1/ Tframe = 1/(5x10-3) = 200 packets / second.
The maximum window size is given by W = 1 + 2(Tprop/Tframe). The RTT = 2Tprop = 0.5. Hence:
W = 1 + (0.5/5x10-3) = 101
(b) Max. Transmission speed is limited by the terrestrial link speed = 1/T tframe = 1/(20x10-3) =
50 packets / second. The maximum window size occurs when the acknowledgment delay is
equal to the time taken to transmit W packets, that is: W T tframe = T tframe + T sframe + RTT
W(20) = 20 + 5 + 500 = 525
W = 26.25 27

(c) For the terrestrial link RTT = 0, and hence W = 1. For the satellite link the time taken to
transmit W packets is limited by the terrestrial link (otherwise the intermediate router buffers
would become infinitely large), but the acknowledgment delay is restricted to the satellite link
W = 25.25 26
only, thus we have: W T tframe = T sframe + RTT and W(20) = 505

2.

Consider a token bucket with token arrival rate of bytes/sec and a token credit capacity
of W bytes. The data rate is M bytes/sec. Let S be the duration of the maximum output
burst rate from the full token-bucket.
(a) Carefully derive an expression for S.
(b) What is S if = 16 Mbps and W = 500 KB and if a burst arrives over a channel
with data rate 200 Mbps and lasts 40 msec?
(c) Complete (b) by indicating what happens after S seconds, assuming bursts arrive
every second.

Solution
(a) Total amount of data in burst = MS. This corresponds to not only the token capacity of the
bucket, W, but the extra tokens that arrive during the burst, S. Thus: W + S = MS, and
hence S = W/(M - ).

(c) The maximum burst output of 200 Mbps (25 MBps) lasts only 22.3 msec which means
(25x106) x (22.3x10-3) = 557.5x103 bytes (544.4KB) of data was delivered by the token
bucket. However the incoming burst lasts 40 msec and this means (200x106 x 40x10-3) = 106
bytes of data was delivered to the bucket, that is 442.5x103 left after 22.3 msec which then
must drain at the token rate of 16 Mbps (2 MBps = 2000 KBps). Thus, there is an additional
16 Mbps burst from the token-bucket after 22.3 msec of (442.5/2000) = 221.25 msec duration.
Hence the token-bucket output exhibits the following cyclic behaviour:
200 Mbps burst of 22.3 msec duration, 16 Mbps burst of 221.25 msec duration and then
756.45 msec of no output.

3.

The standard packet-counting implementation of the leaky bucket algorithm allows

one packet to be transmitted per token generated, independent of the length of the
packet. The byte-counting variation defines a token not as 1 packet but k bytes. Thus if
k = 1024, one token can be used to transmit either one 1024-byte packet, or two 512byte packets (conversely two tokens are needed to transmit one 2048-byte packet). What
are the pros and cons of these two approaches for congestion rate control in networks?

Solution
The byte-counting variation achieves a smoother average bit rate than packet-counting.
However it does not achieve a perfect constant bit rate since tokens will accumulate for large
packets which will then be transmitted at full speed. The byte-counting variant also has the
following problems:
The length of the packet needs to be determined, and a count kept, hence a more complex
implementation.
If packets are small the output packet rate will be high and since routers queue and process
on a per packet basis this may actually exacerbate end-to-end delays and consequent
congestion.
The packet-counting version is simpler to implement and helps control congestion more
effectively since the packet rate is constant, but it also has problems:
A large variance in packet sizes will mean a large variance in bit rate.
Longer packets will be given an effective higher data rate, which may create congestion
problems on networks with smaller MTUs (i.e. packet is segmented as many smaller
packets which will then create longer queues and processing delays).

4.

A leaky bucket is used to control the rate of traffic which arrives as a Poisson
distribution with an average rate of packets per second. The packet lengths are
exponentially distributed with an average length of L bytes. The average token rate is r
bits/sec seconds. What is the average waiting time an arriving packet waits to be
transmitted? How would you modify your answer for a token bucket that can accumulate
up to W tokens in credit?

Solution
Simple! This is an M/M/1 queuing system where the service rate, = r/L and hence:
T
L/r
L
Tq = s =
=
1 1 L / r r L
For a token bucket things get more complicated since if there are any tokens in credit arriving
packets are immediately transmitted, and thus the waiting time would be shortened and this
would depend on the maximum amount of credit , W. If one were to perform the gory details
of the analysis the result would be:
1
1
Tq =
e (W / r )( ) =
e W ((1 / L )( / r ))

( r / L)
(see Bertsekas & Gallager, Data Networks, Prentice-Hall, 1992, page 515, Enjoy!)

5.

Consider a single channel with capacity C bps. The net packet arrivals across all N
stations is Poisson distributed with an average arrival rate of frames/sec and each frame
is drawn from an exponential probability density function with mean L bits/frame.
(a) Derive the mean time spent in the system, T. What convenient assumption has been
made in this analysis?
(b) Now assume FDM is used to allocate N independent identical sub-channels to each
of N stations. Derive the mean time spent in the system using FDM, TFDM. Can you
see why FDM is not a good idea for a LAN system?

Solution
The service rate in bps is (channel rate in bps) / (average length of frame in bits) = C/L
(a) Since lengths are exponentially distributed this is an M/M/1 queuing system from which we
1
1
=
T=
have that:
(C/L)
The convenient assumption is that access to the single channel occurs in a controlled fashion.
(b) With N independent identical sub-channels each sub-channel has a capacity of (C/N) bps
and the mean input rate per sub-channel is (/N), so:
1
1
N
=
=
= NT ,
TFDM =
(C/NL) ( /N) (C/L)
which makes TFDM N times as worse as using a single shared channel.