Performance Improvements of the IEEE 802.

2 LLC Type 2 Protocol

Ernst Bienack
Technische Universitaet Muenchen
Institut fuer Informatik
8000 Muenchen 2
West-Germany
ARPAnet: bienackQinfovax.informatik.tu-muenchen.dbp&de

Abstract Its operation is as follows: All frames transmitted over

This paper investigates the effects on performance of one subnet are processed, but only those whose destina-
three different modifications to the IEEE 802.2 LLC type tion lies within the other partition of the internetwork
2 protocol - further referred to as ORIG - in an internet- are forwarded into the other subnet. In case of overload,
work environment consisting of three IEEE 802.5 token incoming frames that cannot be processed are discarded.
ring networks connected via bridges to a backbone ring.
The modifications are: (1) selective reject (SREJ) of lost The paper is organized as follows. In the next chapter,
or erroneous frames, (2) acknowledgement accumulation the LLC type 2 protocol as standardized by the IEEE
with a static thresho€d value ( A A S ) , and (3) ack- (IEEE 802.21 is briefly described and the elements of an
nowledgement accumulation with a dynamic threshold internetwork are introduced. Chapter III discusses the
value (AAD). problems observed with the original protocol, a
The prime performance measures are throughput and modification to OIUG proposed by [BUX 851, and the
mean end-to-end delay. Additional performance measures three enhancements we make to the original protocol.
such as the percentage of frames lost or being out-of-
sequence are used to gain insight into the internal Chapter IV states the assumptions, defines the perfor-
behavior of the protocols. mance measures, and describes our simulation model.
The modifications (1) and (3) yield a significant improve- Chapter V presents the results gained by running simula-
ment in performance under a high traffic load as com- tions. Chapter VI summarizes our findings.
pared to ORIG. The modification (2) increases the
throughput at the cost of a higher mean end-to-end 11. Framework
delay.
A. LLC Type 2 Protocol
For various types of applications, the IEEE has stand-
I. Introduction
ardized three different layer 2b protocols: The LLC type
For various reasons, it may not be possible to connect all 1, providing a datagram-like connectionless transmission
computers within one organization by a single local area service, the LLC type 2, providing a connection-oriented
network. Among the remons are limited extension or transmission service and the LLC type 3, providing a
diameter, limited number of stations, limited capacity, or guaranteed delivery connectionless transmission service
security considerations. An internetwork consisting of and a poll with guaranteed response [IEEE 802.2, IEEE
several subnetworks connected by gateways can be built 802.2/AD]. To describe the LLC type 2 protocol, we
up to overcome these limitations. introduce the following notation. A packet denotes
The interconnection must take place at one of the seven whatever the layer 3 delivers to layer 2b and a frame is
OS1 layers. The choice of the appropriate layer will be the unit of information sent out by layer 2b. There exist
influenced by the type of networks to be interconnected, three different types of frames: I-frames to carry user
by performance considerations and the protocols used data, S-frames for error control and flow control, and
within the different subnetworks. Above the layer of U-frames to establish and close a connection. N(S)
interconnection, the same protocols must be used denotes the sequence number of the last I-frame
between two communicating entities [BIER 881. In the transmitted by the sender, N ( R ) the sequence number of
configuration considered in this paper, the interconnec- the I-frame expected next by the receiver. [V(S),
tion is achieved by bridges, which are gateways operating V(S)+w-l] and [V(R), V(R)+w-l] denote the current
at sub-layer 2a. Bridges are easy to build and allow to sender and receiver window respectively, with w being
interconnect LANs of different type, e.g CSMA/CD, the window size.
token ring, or token bus LANs [SINC 86). In the follow-
ing, we assume that a bridge interconnects two subnets.

36
CH2613-8/88/0000/0036/$01.000 1988 IEEE
The LLC type 2 protocol has been derived from the layer At one instant of time only the station that has the
2 HDLC protocol for point-to-point links. For a com- token is allowed to transmit frames onto the ring. This
parison of both protocols see [FIEL 861. The LLC type 2 station keeps the token during transmission and passes it
protocol guarantees for an error-free and in-sequence on at the end of the transmission. A station holding the
delivery of all I-frames submitted for transmission. This token is only allowed to transmit one frame at a time. A
is achieved by the use of sequence numbers for I-frames bridge acts as a bidirectional store-and-forward device. It
and positive ( R R ) and negative ( R E 4 acknowledge- receives all frames transmitted over the two rings it is
ments: The receiver generates a positive acknowledge- connected to and transmits (1) those frames received
ment for an I-frame that is received free-of-error and in- from the local ring whose destination lies outside of this
sequence. In case that an I-frame is out-of-sequence or ring onto the backbone and (2) the frames received from
garbled, the receiver discards this frame and generates a the backbone ring whose destination lies within the local
negative acknowledgement (REJ-frame), requesting the ring the bridge is connected to onto the local ring. To
sender to transmit the I-frame with sequence number store the frames before they can be forwarded, the bridge
N(R) next. In the meantime, the receiver drops all I- has a limited amount of buffer space available for either
frames with sequence numbers greater than N(R), until it direction. If the buffer space is exhausted on arrival of a
receives the I-frame with sequence number N(R). To new frame, this frame is simply discarded. A more
provide for the loss of acknowledgements, the sender sets detailed description of the components of the internet-
a timeout (ock-timer) for each I-frame transmitted. Upon work considered in our study is given in [BUX 851.
expiry of this timer, the sender transmits a RR-frame
with the P-bit set to one, to request the acknowledge- 111. LLC Type 2 Protocols being investigated
ment of the I-frames received by the receiver (check-
pointing). The receiver then responds by generating a A. Statement of the Problem
RR-frame with the F-bit set to one, acknowledging the Bux and Grillo performed [BUX 851 a detailed analysis
I-frames received in-sequence. To control the flow of I- of the LLC type 2 protocol in an internetwork environ-
frames, the LLC type 2 protocol uses a sliding window ment. Their study revealed a severe performance degra-
mechanism with a fixed window size. The receiver also dation under high load. The authors concluded that the
can stop the sender by issuing a RNR-frame. flow control scheme keeping the window size fixed is
inadequate. They suggested a modification to the LLC
B. Elements of t h e Internetwork type 2 protocol using a dynamic window mechanism.
The internetwork consists of three token rings connected The proper choice of the right window size has a great
via bridges to a backbone ring as depicted in figure 1. impact on the behavior of the LLC type 2 protocol under
various load conditions and traffic patterns. While a
window size of w=10 yields good throughput and delay
characteristics under moderate load, a severe throughput
degradation can be observed if the'load is raised beyond
a certain level. Using a small window size of w=l or
w=2 might help in this case, but it is not always the
best choice: Under moderate load, the mean end-to-end
delay for small window sizes is considerably higher than
for larger ones. This observation led Bux and Grillo to
propose a dynamic flow control scheme adapting the win-
dow size to the different load conditions [BUX 851. The
details of this scheme will be discussed in section V.B.
Unfortunately, this scheme can not be realized within the
scope of the LLC type 2 protocol (ORIG) as standard-
ized by the IEEE (IEEE 802.21.
We investigate the effects on performance of three other
modifications of ORIG. One modification is concerning
Figure 1: Structure of the internetwork the introduction of a SREJ-frame, the other two explore
different strategies for generating positive acknowledge-
The IEEE standard tor toaen nngs provides for eight ments. The last two are variations of ORIG protocol
levels of transmission priority [DIXO 87). In our study, and are implementable within the scope of ORIG as
we consider only the case that all stations operate at the standardized by the IEEE.
same level of priority. The layer 2a protocol of a token
ring uses a conflict-free and deterministic access scheme.

37
B. Selective Reject (SREJ) IV. Experimental Method
To request the retransmission of an I-frame, the ORIG A. Assumptions
protocol generates a REJ-frame with N(R) being the
sequence number of the I-frame that is missing or gar- The assumptions listed in this section are identical to
bled. The receipt of the REJ-frame causes the sender to those made in [BUX 851. This helps us validate our simu-
re-transmit all I-frames with sequence numbers out of lator and allows us to compare our results with the ones
[N(R), N(S)]. The receiver discards all I-frames with presented in [BUX 851.
sequence numbers greater than N( R) that are received The transmission rate for the local ring and backbone
before the expected I-frame arrives, with the sequence ring is 4 Mbps each. The bridge buffer size for either
number being equal to N(R). Unlike the REJ-frame, a direction is 4 Kbyte. The frame-processing within a
SREJ-frame with sequence number N(R) calls only for bridge takes 300 ps. The flow control between the LLC
the retransmission of the I-frame with the sequence type 2 layer and the higher layer is achieved by back-
number equal to N(R). The receiver retains all out-3- pressure with a threshold value set to window size+4.
sequence I-frames with sequence numbers out of The execution times of the LLC type 2 protocol are
[N(R)+l, V(R)+w-11 that were successfully received chosen as follows: Transmit I-frame (first time): 2 ms;
before the I-frame with the sequence number N(R). receive in-sequence I-frame and transmit RR-frame: 2.5
ms; receive RR-frame and remove acknowledged I-
The use of the SREJ-frame is appealing, because the frames: 0.75 ms; receive out-of-sequence I-frame and gen-
sender can save transmission time and bandwidth by erate REJ-frame: 1 ms; receive REJ-frame or RR-frame
skipping over those subsequent I-frames that successfully
with F-bit set to one and retransmit I-frames: 1 ms;
got to the receiver before the I-frame whose retransmis-
receive and discard out-of-sequence I-frame: 0.5 ms;
sion was requested. Since the receiver always has to pro- receive RR-frame with P-bit set to one and transmit
vide sufficient buffer space for each connection to store RR-frame with F-bit set to one: 1 ms; handle interrupt
up to w frames, the SREJ scheme does not need more timer and transmit RR-frame: 1 ms.
buffer space than ORIG.
The value for the ack-timer is chosen to be 250 ms. The
C. Acknowledgement Strategies interarrival times of packets are exponentially distri-
buted. The packet length is hyperexponentially distri-
The specification of the ORIG protocol leaves to the
buted with a mean of 1 Kbyte and a coefficient of varia-
implementor the choice to acknowledge an I-frame by (1)
tion of 1.5. For a motivation of this particular choice see
immediately sending a RR-frame or (2) waiting - if this
[BUX 851. The length of a S-frame is 24 bytes, the max-
I-frame does not have the P-bit set to one - for further
imum length of an I-frame is 512 byte.
I-frames to arrive and to acknowledge them altogether
by a single acknowledgement. The later strategy is We consider a configuration where each local ring con-
referred to as acknowledgement accumulation. The inter- sists of 12 stations, each 6 of them exchanging data with
val of time the receiver may defer the generation of an 6 stations in one of the two other local ring networks, i.
acknowledgement has to be bounded by a threshold value e. there is only internet traffic with the transmission
th such that unnecessary retransmissions due to delayed being full duplex. Following [BUX 851, we do not use
acknowledgements are avoided. The threshold value th the RNR-frame for flow control at layer 2b and imple-
can be chosen to be either static or dynamic, e.g. ment the ORIG protocol such that a RR-frame is gen-
depending on the current traffic load. See sections V.D. erated for each I-frame received in-sequence and free-of-
and V.E. for a more detailed description. error.
The two acknowledgement strategies investigated in the
B. Performance measures
following are (i) accumulation of acknowledgements with
the threshold value being static and (ii) accumulation of The two prime performance measures are throughput and
acknowledgements with the threshold value being mean end-to-end delay. Both are measured at the inter-
dynamic. face between the LLC type 2 protocol and the layer
above referred to as higher layer. Throughput is defined
Before we discuss the effects on performance of the vari-
as the number of bits per second delivered by the LLC
ous modifications of ORIG, we state the assumptions
type 2 protocol to the higher layer. The mean end-to-
made, define the performance measures, and briefly
end delay is defined as the average time between supply-
describe the simulator.
ing a packet to the LLC type 2 layer at the source sta-
tion and its complete delivery to the higher layer at the
receiving end.

38
We also use the following performance measures. The Figure 3 presents an explanation for the performance
percentage of out-of-sequence I-frames, i. e. the I-frames degradation. As the internet becomes loaded, the buffer
received and discarded by the LLC type 2 layer because space at the bridges fills up and the loss rate soars up.
they were out-of-sequence and the percentage of I-frames
dropped by the bridge because of buffer shortage. The
probability density function of the window size. This is 0RIG
only meaningful if the window size is dynamic. The pro-
...... out-of-sequence
80 - - lost ....'...w=10.
bability density function of the gap size between two suc-
cessive I-frames that have been successfully received. - 1
E 60

C. Measurement and Analysis of the Data

To observe the effects of the modifications made to the
LLC type 2 protocol, we have written a simulator in
Simula 67 using a process-oriented approach to system
simulation. The entire system is represented by a set of
processes interacting with each other [FISH 781. During 0
I I I I
dotal ofiered &a r a i [ M b i ]
one simulation run numerous data concerning the state
of the system are gathered. Since the data normally can- Figure 3: Percentage of out-of-sequence
not be assumed to be independent, we use the batch I-frames and percentage of I-frames lost
means technique with a variable batch size to obtain a by bridge versus total offered data rate
set Xi of independent observations of a random variable
X (e.g end-to-end delay or throughput) [LAVE 831. We The difference in loss rate between w=5 and w=10 is due
approximate the expectation p of X by its mean value to the fact that for greater window sizes the number of
E(X). The simulation stops if pr(E(X)-1/10*E(X) 5 p frames to be retransmitted is also greater, thus making
5 E(X)+l/lO*E(X)) = 0.95 for the performance meas- the congestion - due to an overloading of the bridge -
ures mean end-to-end delay and throughput [FISH 781. worse. Figure 3 also shows that in the case of congestion
a high percentage of I-frames getting through to its desti-
V. Results nation is out-of-sequence and has to be discarded. This
In order to set up the proper context for the evaluation leads to a high number of retransmissions per I-frame.
of our enhancements, we first present some results con- For w=l, there can never arrive an I-frame out-of-
cerning the performance of ORIG and DYN. The figures sequence, since the transmission of the next I-frame does
2, 4 and 5 depict results also presented in [BUX 851. not happen before the preceding one has been success-
fully transmitted and acknowledged.
A. LLC Type 2 Protocol (ORIG) As far as the throughput is concerned, a window size
The following three figures illustrate the problem with w=l yields the best performance. However, due to its
the IEEE LLC type 2 protocol. In figure 2, a severe stop-and-wait behavior, w = l does not allow the sender
throughput degradation can be observed for window sizes to make full use of the available bandwidth under
w=5 and w=10 at a total offered data rate greater than moderate load. For moderate load conditions, w = l
2 Mbps, which is due to the congestion of the bridge sta- results in a higher end-to-end delay than w=5 and w=10
tions. - see figure 4.
4
-rA
2000 iw=lU
OFUG
a 1000
-
$ 3
\ w=5
Y

a 250 \
. e 2 \

z
2
5 1
d
e............................

d w=10
Y

g o I I I I I
i 3 4 5 0 1 2 3 4 5 6
Total ozered data rate [Mbps] Total throughput [Mbps]

Figure 2: Total throughput versus Figure 4: Mean end-to-end delay versus

total offered data rate total throughput

39
B. Dynamic Window (DYN)
The observation of the problems with ORIG led Bux and
Grillo to propose a dynamic flow control scheme - here 0.6 initial window iw=lO
referred to as DYN - [BUX 851, which adapts the win-
dow size to the current traffic load.
At the beginning, the sender initializes its window size w
with the value iw, with iw being the initial window size
that takes a large value such as 10. The sender switches
to a window size w=l as soon as an I-frame has to be
retransmitted. This scheme prevents the bridge form -8
I

I
t--1

I I I I I
- ___
-__-_
I I I
being overloaded by retransmissions and transmissions of 0 1 2 3 4 5 6 7 8 9 10
Window size
new I-frames. In the following, as long as w < iw holds,
the window size w is increased by one after n successful Figure 6: Probability density of
transmissions in a row, with n being a parameter to be the dynamic window size
tuned. The greater the value of n, the longer it takes
until w will be incremented by one.
Despite its merits, the DYN scheme suffers from the
Figure 5 plots the throughput for several values of n. disadvantage that in order to integrate it into the IEEE
One can see that for larger values of n the throughput 802.2 LLC type 2 standard, a revision of the standard is
behavior of DYN under higher load does not degrade and necessary. In the following 3 sections of this chapter, we
is close to the one for ORIG with w=l. A good perfor- will investigate other ways of improving the performance
mance can also be observed for the mean end-to-end of ORIG. Among others, we propose a variation of
delay - see figure 11. According to [BUX 851, we will ORIG which can be implemented without modifying the
always choose the initial window size iw to be 10 and set IEEE 802.2 LLC type 2 standard.
n to 8 (DYN n=8).
C. Selective Reject (SREJ)
The first modification we make, is to introduce another

3 1 S-frame, the SREJ-frame, allowing the receiver to

request the retransmission of exactly one I-frame. This
is motivated by the observation that the use of the

2l A
REJ-frame leads to a go-bock-N behavior, starting the
retransmission with the first I-frame that was not
received properly. Subsequent I-frames, no matter
1
static window w=ld whether they previously got to the receiver or not, are
retransmitted. This aggravates the congestion at the
0 I I I I I bridge and causes a throughput degradation as depicted
0 1 2 3 4 5 6 in figure 2. The usage of the SREJ prevents unnecessary
Total offered data rate [Mbps] retransmissions. Figure 7 shows that for ORIG with
w=10, even under a very high load, in more than 50% of
Figure 5: Total throughput versus the time the number of consecutive I-frames lost is just
total offered data rate one. This supports the hypothesis that the retransmis-
sion of many I-frames will be unnecessary if the I-frames
The DYN scheme prevents the bridge from being over- arriving out-of-sequence are retained to be reordered
loaded and achieves good performance for the various later, after the missing I-frame(s) is (are) received.
traffic conditions. The number of I-frames lost at the
bridge never exceeds 7%, (figure 12) as does the number
of I-frames arriving out-of-sequence at the destination.
Figure 6 shows how the DYN scheme adapts the window
size to the load conditions. Under a moderate load, the
window size stays more than 80% of the time at w=10,
allowing the sender to make full use of the available
bandwidth. Under overload, for about 55% of the time,
the window size stays at w=l, preventing the bridge
from being overloaded.

40
 Although the performance improvements 9f the DYN
and the SREJ scheme are almost identical, the underly-
ing behavior of the two protocols is quite different.
ORIG
x window w=10 Let's consider the case that an I-frame gets lost and that
Total offered data rate: the loss is detected by the receiver. In case of the DYN
- 2Mbps scheme, the receiver generates a REJ-frame and the
P - - 6 Mbps
2 sender reduces the window size to w=l, i.e. it enters a
0.2 stopand-wait mode. This results in choking the
-- I retransmission of I-frames as well as the transmission of
n
"
_-_ new I-frames. In the case of the SREJ scheme, the
I I I I I I I I
0 1 2 3 GQaps~e 6 7 8 9
receiver generates a SREJ-frame. The sender responds to
this with the retransmission of a single I-frame and keeps
on transmitting new I-frames with sequence numbers out
Figure 7: Probability density of of [N(S), V(S)+w-11, i.e. it does not choke the transmis-
the gap size in case of sion of new I-frames. As a consequence, the drop rate of
I-frame low I-frames at the bridge under overload is higher - almost
twice as high - for the SREJ scheme than for the DYN
The figures 8 and 9 show that the SREJ scheme achieves
scheme (figure 12). The SREJ scheme compensates for
a throughput and mean end-to-end delay behavior that is
the higher drop rate at the bridge by retaining out-of-
equal or better than the one achieved for the DYN
sequence I-frames at the destination for further reorder-
scheme.
ing.
4 We investigate two more ways improving the perfor-
mance of ORIG. Both exploit the possibility to accumu-
late acknowledgements. As opposed to the DYN scheme
and to the introduction of a SREJ-frame, no
modification to OFUG is necessary. Only the error con-
trol of ORIG is implemented in slightly different ways.
...... AAD 100/10/25 D. Acknowledgement Accumulation with a Static
- - - SREJ Threshold Value (AAS)
- DYNn=8
The accumulation of acknowledgements helps save pro-
cessing time at the stations and, to a marginal degree,
0 1 2 3 4 5 6 save bandwidth by reducing the number of acknowledge-
Total offered data rate [Mbps]
ments generated.
Figure 8: Total throughput versus When accumulating acknowledgements over a certain
total offered data rate interval th, one must make sure that the sender gets the
acknowledgement - either positive or negative - before its
-3 1000 ack-timer expires and triggers the generation of a RR-
frame with P-bit set to one. Otherwise the savings will
I

500 (initial) window = 10

be nullified. In our configuration the ack-timer is set to
4
2 250 250 ms and the value of th is either 50 ms or 150 ms.
a The following two figures 10 and 11 show the effect of

2x
GI 100 the AAS scheme on throughput and mean end-to-end
50 delay. The shape of the curve for the AAS scheme in
25 ***.*. AAD 100/10/2 figure 10 is basically the same as observed for the ORIG
- - - SREJ protocol. The AAS scheme makes the degradation of the
2 - DYN n=8 throughput happen for a slightly higher load and yields a
E 10
minor overall increase in throughput for higher values of
0 1 2 3 4 5 6 the total offered data rate. Since the AAS scheme delays
Total throughput [M bps]
the generation of acknowledgements, the available
Figure 9: Mean end-to-end delay versus bandwidth can not be fully exploited. This results - see
total throughput figure 11 - in higher values for the mean end-to-end
delay as compared to ORIG.

41
 E. Acknowledgement Accumulation with a
Dynamic Threshold Value (AAD)
TO improve the performance of the acknowledgement
I accumulation scheme, we choose to make the threshold
value dynamically adapting to the current load (AAD
scheme).

Let th denote the current threshold value. The AAD

scheme works as follows. The generation of an ack-
nowledgement is deferred up to th units of time. All 1-
frames arriving during this interval of time are ack-
0 1 2 3 4 5 6 nowledged by a single RR-frame.
Total offered data rate [Mbps]
The adaptation works as follows: (1) The initial value of
Figure 10: Total throughput versus th is 0. As long as no loss is observed at the receiver's
total offered data rate side, an I-frame will be acknowledged immediately. (2)
If the receiver detects an erroneous or lost I-frame, a
REJ-frame is generated and the value of t h is set to a
-3 2000 (initial) window = 10
...... DYN n=8 maximum value maa-th. Delaying acknowledgements
- 1000 - - - AAS helps slow down the speed at which new I-hames are
.$ 500 - ORIG submitted into the system. (3) After a number I of I-
250 frames has been successfully received in a row and if th is
a greater than 0, the value of th is reduced by a fixed
100
value red-th. The values for max-th, red-th, and z have
2 50 to be chosen properly. Our experiments over a range of
25 values suggested the following choice:
max-th=lOO ms, z=10, red_th=25 ms (AAD
100/10/25).
0 1 2 3 4 5 6
Total throughput [Mbps] The performance figures for the AAD scheme compare
favorably to those of the DYN and SREJ scheme. Fig-
Figure 11: Mean end-to-end delay versus ure 8 plots the throughput characteristics for the AAD
total throughput scheme. Its behavior is close to optimal. For a total
offered data rate in the range between 2.5 and 4.0 Mbps,
As can be seen from figure 12, the high level of loss at the throughput is slightly lower than for the DYN
the bridge remains the same as for OFLIG, only going to scheme. The mean end-to-end delay - see figure 9 - in
happen at a slightly higher load. Therefore, the problem this range is slightly higher for the AAD scheme than for
of overloading the bridge is not solved by the AAS the DYN scheme. The AAD scheme chokes the submis-
scheme. sion of new I-frames more drastically than the DYN

I LT/
scheme. This results in an extremely low loss rate at the
100 bridge being always less than 5% - compare figure 12 -
I (initial) window = 10
- - SREJ and in a very low rate of I-frames arriving out-of-
- DYN n=8 sequence.
.... AAD 100/10/25
Figure 13 plots the 95% percentile of the end-to-end
delay for the AAD, SREJ, and DYN scheme. We can
- _ - - - - - - I
40 see a close correlation between the loss rate of I-frames
(figure 12) and the value of the 95% percentile. The
20 scheme with lowest loss rate achieves the lowest 95%
1
-------- percentile. For a total offered data rate greater than 3
0 .............................
-0-

Mbps, the 95% percentiles of the DYN and SREJ scheme

0 1 2 3 4 5 6 are nearly two times higher than for the AAD scheme.
Total offered data rate [Mbps]

Figure 12: Percentage of I-fames lost by

bridge versus total offered data rate

42
 References
1000
1 (initial) window = 10
7
L
500 [BIER881
Biersack, Ernst: ”Techniken zum Zusammenschlufl von
Q, 250 Rechnernetz6 und deren Anwendung auf Protokolle des
rY
;:
Transportsystems” . Dissertation T U Miinchen, February
E!2 100 1988.
& 50
[BUX851
%
0)
25 ...... A A D inn Bux, Werner; Grillo, Davide: ”Flow control of intercon-
nected token rings”. IEEE Transactions on Communica-
10 tions, vol COM-33, no 10, pages 1058-1065, October
I I I I I I
1985.
0 1 2 3 4 5 6
Total throughput [Mbps] [DIXO 871
Dixon, Roy C.: ”Lore of the token ring”. IEEE Network
Figure 13: 95% Percentile of the end-to-end Magazine, vol 1,no 1, pages 25-32, January 1987.
delay versus total throughput
[FISH781
A similar correlation was observed between the standard Fishman, George S.: ”Principles of discrete event simula-
deviation of the end-to-end delay and the loss rate. tion”. John Wiley & Sons, New York 1978.
[GERA 871
VI. Conclusion Gerauer, Bernhard: ”Simulation eines Multiringnetzes” .
To make optimal use of the available bandwidth, the Diplomarbeit TU Miinchen, November 1987.
LLC type 2 protocol has to strike the balance between [IEEE 802.2/AD]
keeping the loss rate sufficiently low to make the waste Institute of Electrical and Electronics Engineers: ”Draft
of bandwidth tolerable, and not to be overly cautious by proposed addendum to IEEE 802.2 logical link control,
restricting the submission of I-frames to much. acknowledged connectionless service”. Nov. 1985.
We have been able to show that this can be achieved in [IEEE802.21
several ways. One is the introduction of a SREJ-frame, Institute of Electrical and Electronics Engineers: ”Logical
another the use of acknowledgement accumulation with a link control”. IEEE Standard 802.2-1985.
dynamic threshold value. Compared to the DYN scheme [LAVE 831
proposed by Bux and Grillo, the SREJ and AAD Lavenberg, Stephen S. (ed.): ”Computer performance
schemes behave equally well and achieve near optimal modeling handbook”. Academic Press, New York 1983.
performance. Moreover, the AAD scheme has the advan-
tage being implementable without modifying the stand- [SINC 861
Sincoskie, W. D.: ”Transparent interconnection of broad-
ardized IEEE LLC type 2 protocol. This is achieved by
cast networks”. Proceedings International Zuerich Sem-
making a judicious implementation choice concerning the
inar in Digital Communications, pages 131-134, 1986.
generation of acknowledgements.

Acknowledgement
I would like to thank B. Gerauer [GERA 871 for his help
implementing the simulator and performing the simula-
tions. Discussions with H . 4 . Siegert helped improve the
presentation.

43