A property of MVG_OMALLOOR
1 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002
Optical CDMA with Optical Orthogonal Code
SANGWOOK HAN
shan4@ece.utexas.edu
Department of Electrical and Computer Engineering
The University of Texas at Austin
Austin, TX 78712
Abstract - This report examines optical CDMA communi-
cation techniques with optical orthogonal codes. Simulations
that show the desired properties of theses codes and their use
in optical CDMA are reported. Based on the simulations, we
investigate the properties of optical CDMA. Proba bility of
error is also evaluated.
I. INTRODUCTION
There have been many efforts to take the full advantage
of fiber-optic signal processing techniques to establish an
all optical CDMA communication systems since CDMA
was first applied to the optical domain in the mid-1980s
by Prucnal, Salehi, and others [1-3]. Traditional fiber
optic communication systems use either TDMA or
WDMA schemes to allocate bandwidth among multiple
users. Unfortunately, both present significant drawbacks
in local area systems requiring large numbers of users.
In a TDMA system, the total system throughput is
limited by the product of the number of users and their
respective transmission rates since only one user can
transmit at a time. For instance, if 100 users wish to
transmit at 1 gigabit per second, at a minimum the
communication hardware would need to be capable of
sustaining a throughput of 100 gigabits per second, a data
rate that would strain even the highest performance
optical networking equipment. In addition, TDMA
systems show significant latency penalties because of the
coordination required to coordinate and grant requests for
time slots from users by the central node [4].
Unlike TDMA, a WDMA system allows each user to
transmit at the peak speed of the network hardware since
each channel is transmitted on a single wavelength of
light. A WDMA system could easily support a bandwidth
of one terabit per second, ideal for the needs of a local
area network. Unfortunately, it is difficult to construct a
WDMA system for a dynamic set of multiple users
because of the significant amount of coordination among
the nodes required for successful operation. To build a
WDMA network with a dynamic user base, control
channels and collision detection schemes would need to
be implemented that would waste significant bandwidth.
Fortunately, an alternative to TDMA and WDMA
networking schemes, optical CDMA communication
systems , require neither the time nor the frequency
management systems. Optical CDMA can operate
PDF processed with CutePDF evaluation edition www.CutePDF.com
asynchronously, without centralized control, and it does
not suffer from packet collisions. As a result, optical
CDMA systems have lower latencies than TDMA or
WDMA. Furthermore, since time and frequency (or
wavelength) slots do not need to be allocated to each
individual user, significant performance gains can be
achieved through multiplexing. Also, TDMA and
WDMA systems are limited by hardware because of the
slot allocation requirements. In contrast, CDMA systems
are only limited the tolerated bit error rate relationship to
the number of users, affording the designer a much more
flexible network design [4].
To establish the optical CDMA, we have to overcome
the code orthogonality problem. Many researchers have
proposed several codes such as prime code, optical
orthogonal code, and so on. In this project, we focus on
optical orthogonal codes (OOC) among those codes. In
section II, we introduce the optical orthogonal codes.
Section III discusses three simulations demonstrating the
principles of optical CDMA. The first one is for two-user
synchronous channels to understand basics of optical
CDMA and the second one is for two-user asynchronous
channels. The third one is for K-user synchronous
channels. Section IV evaluates the probability of error.
II. O PTICAL O RTHOGONAL CODES
An optical orthogonal code is a family of (0, 1)
sequences with good auto- and cross-correlation
properties. Thumbtack-shaped auto-correlation enables
the effective detection of the desired signal (Fig. 2 c), and
low-profiled cross-correlation makes it easy to reduce
interference due to other users and channel noise (Fig. 2
d). The use of optical orthogonal codes enables a large
number of asynchronous users to transmit information
efficiently and reliably. The lack of a network
synchronization requirement enhances the flexibility of
the system. The codes considered here consist of truly (0,
1) sequences (Fig. 1 a) and are intended for “unipolar”
environments that have no negative components since you
either have light, or you don't, while most document ed
correlation sequences are actually (+1, -1) sequences (Fig.
1 b) intended for systems having both positive and
negative components.
An (n, w, ?a, ?c ) optical orthogonal code C is a family
of (0, 1) sequences of length n and weight w which satisfy
A property of MVG_OMALLOOR

2 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002

the following two properties [5]. other time following (1). Fig. 2 (d) shows the cross-
1) The Auto-Correlation Property: correlation between OOC 1 and OOC 2. It always takes
n−1 one and zero at any time following (2). Here 1s are taken
∑x xt=0
t t +τ
≤ λa (1) as auto-correlation constraint, ?a, and cross-correlation
constraints, ?c, since 1s are the lowest value they can be,
for any x ∈ C and any integer t, 0<t<n. and correlation is calculated by convolution sum of two
2) The Cross-Correlation Property: sequences.
n −1

∑x y ≤ λc
5 5
(2)
t t +τ 4.5 4.5

t=0 4 4

for any x ? y∈ C and any integer t.

3.5 3.5

3 3

The numbers ?a and ?c are called the auto- and cross- 2.5 2.5

2
correlation constraints. The (0, 1) sequences of an optical 2

1.5 1.5

orthogonal code are called its codewords. 1 1

0.5 0.5

0 0
5 10 15 20 25 30 5 10 15 20 25 30
1.5 1.5

1 1
(a) (b)
5
5
0.5 0.5
4.5
4.5

4
0 0 4
3.5
3.5
-0.5 -0.5 3
3

2.5
2.5
-1 -1
2
2

1.5
-1.5 -1.5 1.5
0 5 10 15 20 25 30 0 2 4 6 8 10 12 14 16
1 1

(a) (b) 0.5

0
0.5

10 20 30 40 50 60 0

Fig. 1. Examples of sequences (a) sequence for fiber 10 20 30 40 50 60

optics (b) sequence for radio frequency (c) (d)

An (n, w, ?a, ?c ) OOC C can be alternatively considered Fig. 2. (a) OOC 1 (b) OOC 2 (c) auto-correlation (d)
as a family of w sets of integers modulo n in which each cross-correlation
w set corresponds to a codeword and the integers within
each w set specify the nonzero bits. For instance, let’s II. O PTICAL CDMA
think of a simple OOC, 1101000, characterized by (7, 3, 1,
II.1.A. Two-User Synchronous Channel
1). Here we can see three 1s as the nonzero bits. Their
positions are 0th, 1st, and 3rd, respectively. Thus 1101000 The two-user synchronous channel simulated here can
can be simply represented by {0, 1, 3} (mod 7). This be characterized by the follows.
notation can simply represent codes instead of
exhaustively describing long (0, 1) sequences. Table I y (t ) = A1b1 s1 (t ) + A2 b2 s2 (t ) + σn( t) A1 = A2 = 1, t ∈ [0, T ] .
(3)
shows some optimal (n, 3, 1, 1) codes in this notation. T  w, s1 = s2
ρ =< s1 , s 2 >= ∫ s1 (t ) s2 (t )dt = 
n Optimal (n, 3, 1, 1)-codes
0
0 or 1, s1 ≠ s2
7 {0,1,3}
Let an (32, 4, 1, 1) optical code C with 2 codewords be
13 {0,1,4}, {0,2,7}
used. C can be represented by {{0, 10, 13, 28}, {0, 5, 12,
19 {0,1,5}, {0,2,8}, {0,3,10}
31}} mod(32) in Fig. 3 (d) and (e). Thus, the system can
25 {0,1,6}, {0,2,9}, {0,3,11}, {0,4,13}
31 {0,1,7}, {0,2,11}, {0,3,15}, {0,4,14}, {0,5,13} accommodate 2 transmitters simultaneously. Each
37 {0,1,11}, {0,2,9}, {0,3,17}, {0,4,12}, {0,5,18}, transmitter is assigned a w(=4) set from C, i.e.
{0,6,12} transmitter1 is assigned a {0, 10, 13, 28} set and
transmitter 2 is assigned a {0, 5, 12, 31} set. At a
43 {0,1,19}, {0,2,22}, {0,3,15}, {0,4,13}, {0,5,16},
transmitter, every information bit is encoded into a frame
{0,6,14},{0,7,17}
of n(=32) optical chips in the following way. (A chip is an
optical time slot which can assume one of two values: ON
Table I. Some Optimal (n, 3, 1, 1) Codes [5]
of OFF) Let the assigned w set for a particular transmitter
be S={s1, s 2, · · · , s w}. In this case, s 1 is a {0, 10, 13, 28}
Let’s think of C represented by {{0, 10, 13, 28}, {0, 5,
12, 31}} (mod 32) with two code words. Two code set and s 2 is a {0, 5, 12, 31} set. Assume the information
bit is 1. In the corresponding frame, which consists of n
words are shown in Fig. 2 (a) and (b), respectively. C also
optical chips, photon pulses (i.e., ON signals) are sent at
can be represented by (32, 4, 1, 1) in (n, w, ?a, ?c ) notation.
exactly the s 1th, s 2th, · · · and s wth chips (Fig. 3. f and g).
Fig. 2 (c) shows auto-correlation of OOC 1. Its maximum
In the other n-w chips, no photon pulse (i.e., OFF signals).
value is w at the correlation time and one or zero at any
A property of MVG_OMALLOOR

3 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002

Transmitting Part Receiving Part

OOC 1 OOC 1
C
A Optical E Optical H J Recovered
+
User 1 • Encoder
• G G Decoder • + • Signal
Optical
Sum • • Optical + Recovered
User 2 •B Encoder • Decoder
• +
•K Signal
F I
I
D
OOC 2 OOC 2
OPTICAL OPTICAL
(a) (a)
5 5 5 5

4.5 4.5 4.5 4.5

4 4
4 4

3.5 3.5 3.5 3.5

3 3
3 3

2.5 2.5
2.5 2.5
2 2
2 2
1.5 1.5
1.5 1.5
1 1
1 1
0.5 0.5
0.5 0.5
0 0
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 0 0
50 100 150 200 250 300 350 50 100 150 200 250 300 350

(b) (c) (b) (c)

5
5
5 5
4.5 4.5
4.5 4.5
4
4
4 4
3.5 3.5
3.5 3.5
3
3
3 3
2.5 2.5
2.5 2.5
2 2
2 2
1.5 1.5
1.5 1.5
1 1
1 1
0.5 0.5
0.5 0.5
0 0
5 10 15 20 25 30 5 10 15 20 25 30 0 0
50 100 150 200 250 300 50 100 150 200 250 300

(d) (e) (d) (e)

5 5

4.5 4.5

4 4

3.5 3.5

3 3 Fig. 4. (a) schematic diagram for a receiving part (b)

2.5 2.5
signal at H (c) signal at I (d) signal at J (e) signal at K
2 2

1.5 1.5

0.5
1

0.5
Basically, optimal CDMA scheme is same as radio
0
50 100 150 200 250 300
0
50 100 150 200 250 300
frequency CDMA scheme except using the special codes.
Here, we use the matched filter to convert the received
(f) (g)
signal in Fig. 3 (h) assuming n(t) in (3) is zero. At the
5

4.5 receiving end, correlation-type decoders are used to

4

3.5
separate the transmitted signals. The decoder consists of a
3 bank of 2 tapped delay -lines, one for each codeword. The
2.5

2
delay taps on a particular line exactly mat ch the signature
1.5
sequence, i.e., the delays between successive taps are
1

0.5
equal to s2-s 1, s 3-s 2, · · · , optical chips, respectively. Each
0
50 100 150 200 250 300 tapped delay-line effectively calculates the correlation of
(h) the received waveform with its signature sequences. In
Fig. 4 (b) and (c), there are five different correlation
Fig. 3. (a) schematic diagram for a transmitting part (b) values, 0, 1, 2, 4, and 5. Because of the properties of
signal at A (c) signal at B (d) signal at C (e) signal at D optical orthogonal codes, the correlation between
(f) signal at E (g) signal at F (h) signal at G different signature sequences is low, 0 and 1. Thus the
delay-line output is high, 4 and 5, only when the intended
On the other hand, if the information bit is 0, no photon transmitter’s information bit is 1. However, a potential
pulses are sent in the corresponding frame, i.e., all OFF problem due to interference can be happened. When
signals are sent. In other words, the codeword set is used correlation value has 2, it is definitely due to the
as the signature sequence of the transmitter. interference. In this case, the value is always below 4 so it
A property of MVG_OMALLOOR

4 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002

can be discarded by choosing relevant threshold value. scheme for synchronization. Therefore, now we can say
But when total number of users goes up, the cross- that optical CDMA does need no network synchronization.
correlation due to interfering users adds up quickly to
severely degrade the system performance. For instance, II.2. K-User Synchronous Channel
when w is 4 like this case, accommodating 4 users make it
possible to have 4 as the correlation value even if the In the previous section, we investigated a simp le 2-user
intended transmitter’s information bit is 0. To avoid this channel to understand the optical CDMA. In this section,
phenomenon, both high w which can be considered as the we explore problems faced by increasing K (Fig. 6 a).
sum of 1s in the sequence, and long n are required. If we Here we choose 7 as K, and C is (43, 3, 1, 1) having 7 sets,
increase only w fixing n, cross-correlation value due to {{0, 1, 19}, {0, 2, 22}, {0, 3, 15}, {0, 4, 13}, {0, 5, 16},
the interference can be lowered. However, OOC has very {0, 6, 14}, {0, 7, 17}}.
sparse marks to keep the cross-correlation low, i.e. a
number of zeros is much higher than that of ones in the The first issue is a threshold value. As we saw in the
sequence. It means that cross-correlation increases by previous section, interfering signal can be effectively
itself by increasing only w. Therefore, both w and n discarded by setting a relevant threshold value. The
should be increased simultaneously. But this solution also threshold value can be chosen under the following
reveals a drawback, long signal processing time due to condition.
long n. 0 ≤ threshold ≤ w (4)

Finally the transmitted information is extracted by Fig. 6 (d) through (g) show results from several threshold
thresholding the correlator output in Fig. 4 (d) and (e). In values, 1, 2, 3, and 5. As the threshold value goes up, the
this case, they are successfully recovered as intended. recovered signal gets similar to the ideal signal in Fig. 6
Here, 3 is chosen as the threshold values. Threshold issue (c). However, the threshold value can not be over w. If the
will be covered in detail in the section II.2. intended information bit is 1, the correlation value is w.
Therefore, the threshold over w incorrectly converts it to
II.1.B. Two-User Asynchronous Channel 0 when the intended information bit is 1. Fig. 6 (g) shows
the results by choosing threshold over w. We can clearly
In optical CDMA, all users are allowed to transmit at see that a dotted line is missed comparing to Fig. 6 (c).
any time. There is no network synchronization required.
In this section, we simulate a two-user asynchronous Even if the highest value under (4) is chosen (Fig. 6 f),
channel to investigate above statement. To verify no the recovered signal can not be exactly same as the ideal
network synchronization, the same scheme in II.I.A. is pattern. In the Fig. 6 (a), if the intended information bit is
used here except the time delay at F in Fig. 3 (a). Fig. 5 1, the correlation value is w(=3). But as we already
(b) and (c) show the recovered signals from asynchronous discussed in the previous section, the cross-correlation
and synchronous channel, respectively. They agree well due to interfering users adds up very quickly to severely
even if synchronous channel does not use any special degrade the system performance. Here, the correlation
value due to interference is even higher than w(=3). This
1.5
is the reason why the received signal can not be recovered
perfectly.
1

To figure it out above problem we introduce optical

0.5 hard-limiter [3] located before the optical tapped-delay
line (Fig. 7 a). An ideal optical hard-limiter is defined as
 1, x≥1 (5)
g ( x) = 
0
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6

 0, 0 ≤ x < 1
(a)
1.5
Therefore, if an optical light intensity (x) is bigger than or
1.5

equal to one, the hard-limiter would clip the intensity

back to one, and if the optical light intensity is smaller
1 1
than one, the response of the optical hard-limiter would
be zero. This ideal nonlinear process would enhance the
0.5 0.5
system performance because it would exclude some
combinations of interference patterns from causing errors
0
50 100 150 200 250 300
0
50 100 150 200 250 300 as in the soft-limiter case, i.e., the patterns that caused
(b) (c) errors by analog summation of light intensity rather than
Fig. 5. (a) Intended information binary signal (b) recoverd by exact reproduction of the particular pattern with no
signal from asynchronous channel (c) recovered signal analog effect. Fig. 7 shows comparison between the
from synchronous channel systems with and without the optical hard-limiter.
A property of MVG_OMALLOOR

5 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002

OOC 1 OOC

Optical H + J Recovered G Optical H J Recovered

Decoder • + • Signal • Decoder • +
+ ? Signal

G . (a)
• .
OOC K .
OOC
Optical + Recovered
Decoder • + • Signal Hard G’ Optical H’ + J’ Recovered
Limiter • Decoder • + ? Signal
(a)
8 1.5 (b)
7
8 8

6
7 7
1
5
6 6

4
5 5

3
4 4
0.5
2
3 3

1
2 2

0 0
0 50 100 150 200 250 300 350 50 100 150 200 250 300 1 1

0 0

(b) (c) 50 100 150 200 250 300 50 100 150 200 250 300

1.5 1.5
(c) (d)
8 8

7 7
1 1

6 6

5 5

0.5 0.5
4 4

3 3

2 2
0 0
50 100 150 200 250 300 50 100 150 200 250 300
1 1

0 0
(d) (e) 0 50 100 150 200 250 300 350 50 100 150 200 250 300

1.5 1.5
(e) (f)
1.5 1.5

1 1

1 1

0.5 0.5

0.5 0.5

0 0
50 100 150 200 250 300 50 100 150 200 250 300

(f) (g) 0
50 100 150 200 250 300
0
50 100 150 200 250 300

(g) (h)
Fig. 6 (a) schematic diagram of a receiving part for K
users (b) signal at H (c) ideal signal at J (d) signal at J 1.5

when threshold=1 (e) signal at J when threshold=2 (f)

signal at J when threshold=3 (g) signal at J when 1

threshold=5
0.5

All optical light intensity which is bigger than or equal

to one are clipped back to one in Fig. 7 (d). In Fig 7 (f),
0
the cross-correlation value over w due to the interference 50 100 150 200 250 300

does not exist any more. From these, the thresolding can
much effectively recover the information. Fig. 7 (g) and
(h) show the recovered signals without and with the
optical hard-limiter, respectively. We can clearly see that
Fig. 7 (h) is much similar to the intended signal (Fig. 7 i).
Consequently, we can say that the optical hard-limiter can
effectively lower the interference effect.
(i)
Fig. 7. (a) receiving part without optical hard-limiter (b)
receiving part with optical hard-limiter (c) signal at G
(d) signal at G’ (e) signal at H (f) signal at H’ (g) signal
at J (h) signal at J’(i) ideal signal expected at J
A property of MVG_OMALLOOR

6 MULTIUSER WIRELESS COMMUNICATION (EE381K) CLASS PROJECT, FALL 2002

IV. P ROBABILITY OF E RROR 10

2

w=1
0 w=3
10
The probability of error per bit is defined as w=5
w=7
-2 w=9
10

10 -4

Probability of Error
PE=p(LI≥th | b=0)?p(b=0)
-6

+p(LI<th | b=1)?p(b=1) (6) 10

10 -8

-10

where PE, LI, th, and b are the probability of error, light 10

-12

intensity, threshold, and intended binary information bit, 10

respectively. Here, an interesting point is that p(LI<th | 10 -14

1 2 3 4 5 6 7 8 9
threshold
b=1) is always equal to zero because p(LI<th |
b=1)=p(w+I < th) = p(w+I-th <0)=0 under (4) where I is (b)
2

the interference due to other users. According to [4], one 10

can easily calculate the probability of error using the

100
followings.

Probability of Error
-2
10

K−1− i
1  K − 1 w2   w2 
i
K −1
(7)
PE = ∑    1 − 
2 i = th  i  2 n  2n 
-4
10


10-6
number of users=10
Fig. 8 shows several dependency of PE. In Fig. 8 (a), a number of users=30
number of users=50
10-8
length of code, n, is tested in different values, 200, 500, 1 1.5 2 2.5 3
threshold
3.5 4 4.5 5

1000, and 2000. As n goes up, PE gets lowered resulting

(c)
in long processing time. So we need to deal with the trade
off problem on the low error rate and processing time. Fig. Fig. 8. (a) n-dependency of probability of error (b) w-
8 (b) shows w-dependency of PE. As the sum of 1s in the dependency of probability of error (c) K-dependency of
code, w, goes up, PE gets lowered. Note that the highest p robability of error
threshold value under (4) would make the lowest PE on
REFERENCES
the same w. In the Fig. 8 (c), we can find K-dependency
of PE. As the number of accommodated users, K, goes up,
PE gets higher. This is definitely due to the increasing [1] P. R. Prucnal, M. A. Santoro, and R. R. Fan,
“Spread Sprectrum Fiber-Optic Local Area
interference. Observations from Fig. 8 agree well with the
Network Using Optical Processing,” IEEE J.
results in the section II.1.A .
Lightware Tech., vol. LT-4, no. 5, May 1986, pp.
V. CONCLUSION 547-54
[2] J. A. Salehi, “Code Division Multiple-Access
In this report, we introduced the optical CDMA with Techniques in Optical Fiber Networks-Part I:
the optical orthogonal codes. CDMA scheme was Fundamental Principles,” IEEE Trans. Commun.,
successfully applied by using the optical orthogonal codes. vol. 37, no. 8, Aug. 1989, pp. 824-33.
However, OOC revealed some drawbacks, requirement of [3] J. A. Salehi and C. A. Brackett, “Code Division
long sequences resulting in long signal processing time Multiple-Access Techniques in Optical Fiber
and severe degradation due to fast adding of cross- Networks-Part II: Systems Performance Analysis,”
correlation. Optical hard-limiter showed remarkable IEEE Trans. Commun., vol. 37, no. 8, Aug. 1989,
improvement in reducing interference due to other users. pp. 834-42.
We also presented that the threshold value, code length, [4] A. Stok, E. H. Sargent, “Lighting the Local Area:
and total number of users are important factors for the Optical Code-Division Multiple Access and
probability of error. Quality of Service Provisioning,” IEEE Network.,
pp . 42-46, Nov.-Dec. 2000.
10
1
[5] F. R. K Chung, J. A. Salehi, V. K. Wei, “Optical
10
0
Orthogonal Codes: Design, Analysis, and
10
-1 Applications,” IEEE Trans. Information Theory,
10-2
vol. 35, no. 3, May 1989.
Probability of Error

10-3
[6] C. F. lam, “To Spread or Not to Spread-the Myths
10-4
of Optical CDMA,” Lasers and Electro-Optics
10
-5
Society 2000 Annual Meeting. LEOS 2000. 13th
-6
n=200
n=500
Annual Meeting. IEEE, vol.2, 2000, pp. 810 -811.
10

-7
n=1000
n=2000 [7] Sergio Verdu. Multiuser Detection. Cambridge
10
1 1.5 2 2.5 3
threshold
3.5 4 4.5 5 University Press, NY, 1998
(a)
A property of MVG_OMALLOOR

Interference Avoidance in Optical CDMA Networks

Purushotham Kamath
Advisors: Joseph D. Touch and Joseph A. Bannister
USC/ISI
Marina del Rey, CA 90292
{pkamath, touch, joseph}@isi.edu

Abstract— Interference Avoidance is a media access control

mechanism that prevents throughput degradation in a broadcast
all optical Local Area Network (LAN) built on optical Code
Division Multiple Access (CDMA) technology. Optical CDMA is
a multiplexing technology that allows the utilization of the large
transmission capacity of an optical fiber. However, the throughput
of an optical CDMA broadcast LAN network degrades quickly
due to multiuser interference. Interference Avoidance controls
interference and prevents throughput degradation due to interfer-
ence at high loads. It is based on two mechanisms: state estimation
and transmission scheduling. Algorithms for state estimation and
transmission scheduling are proposed and evaluated. Analysis
and simulation show how they prevent degradation of throughput
at high loads. A testbed implementation of state estimation and
transmission scheduling hardware is in progress.
Fig. 1. Interference avoidance
I. I NTRODUCTION
Optical CDMA is a code division multiplexing technology
that allows several transmitters to transmit simultaneously on If the receiver is currently receiving a packet, this kind of false
an optical fiber. Nodes transmit using codewords from an Op- positive error can result in the loss of the packet.
tical Orthogonal Codeset (OOC) [1]. An OOC is a set of (0,1) Related work in the area of Optical CDMA mostly focuses
sequences of length N that satisfies certain autocorrelation and on code design [1] or system design [3]. A survey of related
cross-correlation constraints. The term codeset is used to refer work may be found in [2].
to the set of such sequences, and the term codeword is used
This work demonstrates thats Interference Avoidance can
for a member of the set. Each 0 or 1 of a sequence is called
control interference and improve the throughput of an optical
a chip, and the sequence represents a data bit. The number w
CDMA LAN at high loads.
of 1 chips of a codeword of the codeset is called its Hamming
weight. Codesets are designed so that the autocorrelation of
any codeword and the cross-correlation between any pair of II. I NTERFERENCE AVOIDANCE
codewords in the codeset is constrained below a value called
the Maximum Collision Parameter κ. Consider the codewords shown in Figure 1(a). The figure is
A node transmits data by ON-OFF keying (OOK) of a a snapshot of data bits sent by two nodes (codewords 1 and
codeword. Optical CDMA receivers are correlation receivers. 2). A third node is preparing to transmit on codeword 3. If
By connecting nodes on the network in a star, bus or ring transmitted with the chip offset shown in Figure 1(a) a false
topology any node can transmit to any other using the re- positive error would probably occur. If it was sent three chip
ceiver’s codeword. More details on the system architecture of times later (Figure 1(b)), all three packets could be transmitted
an optical CDMA network may be found in [2]. correctly. (Codeword 3 has at least one chip that does not
The advantage of optical CDMA is that it allows utilization interfere with codewords 1 and 2. Hence no false positive can
of the available transmission capacity of an optical fiber. Data occur and it will be received correctly).
processing can occur at electronic speeds while encoding and Interference Avoidance uses the above principle. A node
decoding can be done all optically at a higher chipping rate [3]. estimates the state of the line (state estimation). Given the
The main problem with using optical CDMA in a LAN state estimate and the codeword to be transmitted, it decides
is that at high loads multiuser interference errors result in the appropriate time instant to transmit such that there is no
low network throughput [4]. The reason is that a correlation interference (transmission scheduling). Interference avoidance
receiver set to receive a particular codeword will erroneously reduces interference and improves throughput. Note that any
detect a 1 bit if there are other codewords on the line which packet arrivals in the time between the state estimation and
have ‘1 chips’ in the same locations as the expected codeword. the transmission can still cause interference.
A property of MVG_OMALLOOR
III. S TATE ESTIMATION
Assume that several nodes on an optical CDMA LAN
are transmitting simultaneously on different codewords. Each
codeword is shifted by a different phase shift that depends on
the exact instant it was transmitted. The state of the line at
a point on the line is the sum of the codewords taken over a
window of N chips. E.g., in Figure 1(c), the state of the line
is [0120111].
Arrivals and departures of packets continuously change
this state of the line. The on-off keying of the codewords
also changes the state of the line. The objective of state
estimation is to determine the true state, i.e. the sum of the
codewords being transmitted with appropriate phase shifts. A
window based state estimation algorithm can give a reasonable
estimate of the true state of the line. The node determines the
state by sampling the line. This is called the sensed state This
is repeated for a window of time, say W bits. The estimated
state is twice the average of the sensed states taken over
over the window. State estimation can be either performed
continuously or on-demand, i.e. when a packet arrives.
The performance of an on-demand window based state es-
timation algorithm has been measured through simulation and
the results of system performance are described in Section V.
State estimation hardware needs to be implemented using
sampling over several bits because chips are arriving faster
than the maximum electronic processing rate. The hardware
for state estimation is currently being implemented.
IV. T RANSMISSION SCHEDULING
Given an accurate estimate of the true state of the line, a
node can use a transmission scheduling algorithm to determine
the appropriate time to transmit.
The following classes of transmission scheduling algorithms
have been identified:
• Selfish algorithms: These algorithms schedule the trans-
mission of a packet such that other codewords on the line
will not interfere and cause loss of this packet
• Cooperative algorithms: These algorithms schedule the
transmission of a packet such that the transmission of
this codeword will not interfere and cause loss of other
packets on the line.
• Pseudo-cooperative: These algorithms control interfer-
ence by limiting either the number or size of ’1’ chip
overlaps.
Transmission scheduling algorithms can be analyzed using a
Markov chains [5] approach by reducing the state description
to a simpler form to make it mathematically tractable. The
average throughput of different transmission scheduling algo-
rithms may obtained from this model. The average throughput
of the algorithms may be compared to an upper bound on the
throughput of an optical CDMA network obtained through a
combinatorial analysis [5].
The transmission scheduling algorithms will be imple-
mented in electronics while the transmission delays can be
controlled by selecting from a set of optical delay lines. The
Throughput vs. offered load for different Transmission Scheduling algorithms
1
No transmission scheduling
Pseudo-cooperative
Selfish
0.8 Cooperative
Normalized throughput
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Normalized offered load
Fig. 2. Throughput vs. offered load for different transmission scheduling
algorithms for simulation of a star network using a (10, 3, 2) codeset. A
continuous window based state estimation algorithm was used with sensing
window = 10 bits. Four transmission scheduling algorithms were evaluated:
Selfish, Pseudo-cooperative, Cooperative and None (Aloha-CDMA). A 1-
persistent defer algorithm was used with a retry limit of 10.
transmission scheduling hardware is currently being imple-
mented.
V. S YSTEM PERFORMANCE
The performance of an Interference Avoidance system is
dependent on a large number of factors such as traffic patterns
and network topology. An analysis based on Poisson arrivals,
exponentially distributed packet lengths random codeword
assignment, perfect state estimation may be found in [2].
Figure 2 shows the results of a realistic simulation that uses
a traffic model based on real network traffic. The protocol
is called Interference Sensing/Interference Detection (IS/ID).
Nodes estimate the state and schedule transmissions. If in-
terference is detected during transmission, the transmission
is aborted and the node backs off. The simulation assumed
a star network, a window based on demand state estimation
algorithm, three different types of transmission scheduling
algorithms and a defer algorithm for retransmission [6].
Both analysis and simulation confirm that the throughput
of a system without Interference Avoidance degrades with
increasing load, while the throughput of an Interference Avoid-
ance system remains constant at high loads.
R EFERENCES
[1] H. Chung, J. Salehi, and V. Wei, “Optical orthogonal codes: Design,
analysis, and applications,” IEEE Transactions on Information theory,
vol. 35, no. 3, pp. 595–605, May 1989.
[2] P. Kamath, J. D. Touch, and J. A. Bannister, “The need for media access
control in optical CDMA networks,” in IEEE Infocom, vol. 4, March
2004, pp. 2208–2219.
[3] J. Salehi, “Code division multiple-access techniques in optical fiber
networks - Part 1: Fundamental principles,” IEEE Transactions on Com-
munications, vol. 37, no. 8, pp. 824–833, Aug. 1989.
[4] C. Lam, “To spread or not to spread: The myths of optical CDMA,” in
IEEE Lasers and Electro-Optics Society Annual Meeting, vol. 2, 2000,
pp. 810–811.
[5] P. Kamath, J. D. Touch, and J. A. Bannister, “Transmission scheduling
in optical CDMA networks,” in Under submission, Jan 2005.
[6] ——, “Algorithms for interference sensing in optical CDMA networks,”
in IEEE International Conference on Communications, vol. 3, June 2004,
pp. 1720–1724.
A property of MVG_OMALLOOR
ISI-TR-2006-617
Algorithms for Transmission Scheduling
in Optical CDMA Networks
Purushotham Kamath, Joseph D. Touch, Joseph A. Bannister
USC/ISI
4676 Admiralty Way
Marina del Rey, CA 90292
{pkamath, touch, joseph}@isi.edu
22nd May 2006
Abstract— Transmission scheduling is a media access control
mechanism that prevents degradation of throughput in optical
CDMA Local Area Networks (LANs) at high offered load. Optical
CDMA is a multiple access technique for broadcast optical
Local Area Networks. The throughput of an optical CDMA
LAN at high offered load is limited by multi-user interference.
Interference Avoidance, a distributed, contention based media
access control mechanism, can prevent throughput degradation
at high loads. Interference avoidance consists of state estimation
and transmission scheduling. This work analyzes algorithms for
transmission scheduling under perfect state estimation. The anal-
ysis shows that transmission scheduling under specific conditions
can provide upto 30% network throughput at high offered load.
This compares well to non scheduled systems which have close
to zero throughput under the same conditions. Simulations show
that the performance of transmission scheduling is independent of
codeset length and degrades with increase in codeset weight. The
results also show that the performance of transmission scheduling
does not degrade when used with realistic network traffic based
on traffic obtained from a real network.
KEYWORDS: Networks, Optical communication, Code divi-
sion multiaccess, Protocols, Access control.
I. I NTRODUCTION
This work considers a shared medium, packet switched
optical CDMA LAN in which several nodes are connected
to a passive star coupler to form an all optical broadcast
network. Each node on the network is allocated an optical
CDMA codeword to receive on. Optical CDMA codewords
are sequences of zeroes and ones (unipolar codewords) that
are transmitted asynchronously. The codewords are transmitted
by binary intensity modulation i.e. a one in the codeword is
represented by pulse of light. Nodes use ON-OFF keying of
the codeword to transmit binary data. To transmit a 1 bit the
codeword is sent and to transmit a 0 bit, an all zeros codeword
is sent. When a node wants to transmit, it tunes its transmitter
to the receiver’s codeword and transmits. The code division
multiplexing allows several pairs of users to communicate
simultaneously.
1 This material is based upon work supported by the Defense Advanced
Research Projects Agency under contract no. N66001-02-1-8939 issued by the
Space and Naval Warfare Systems Center (SPAWAR). Any opinions, findings,
and conclusions or recommendations expressed in this material are those of
the authors and do not necessarily reflect the views of the Defense Advanced
Research Projects Agency, SPAWAR, or the U.S. Government.
1
The throughput of an optical CDMA LAN is limited by
multi-user interference. When several users transmit simul-
taneously, their packets and hence their codewords overlap.
When the optical pulses in the codeword overlap, their optical
power is added. Optical pulses from one codeword can be
detected by receivers tuned to other codewords. As a result
receivers may falsely detect their codewords resulting in packet
errors. These false positive errors increase with offered load,
resulting in throughput collapse.
Interference Avoidance is a contention media access control
mechanism that prevents throughput collapse in optical LANs
networks at high offered load. It consists of state estimation
and transmission scheduling. State estimation is a mechanism
by which nodes on the network estimate the state of the line.
Transmission scheduling is a mechanism by which nodes use
the estimated state to schedule their transmissions to avoid
packet losses due to interference.
The contribution of this paper is the analysis of transmission
scheduling algorithms for optical CDMA under perfect state
estimation. The analysis quantifies the difference between
throughput of systems with and without transmission schedul-
ing. The analysis shows that transmission scheduling under
specific conditions can provide upto 30% throughput at high
offered load. In contrast, non scheduled systems have close
to zero throughput under the same conditions. A sensitivity
study of the transmission scheduling algorithms shows that the
performance gain depends on certain codeword parameters.
The performance is independent of codeword length and
degrades with increase in codeword weight. The performance
does not degrade with a traffic model based on traffic obtained
from a real network.
The paper is organized as follows. Section II provides
background on optical CDMA. Section III-A discusses the
motivation for Interference Avoidance. Section III-C discusses
the channel characteristics of optical CDMA and a represen-
tation for the state of the network. Section III-D defines the
transmission scheduling algorithms. Section IV analyzes the
performance of the transmission scheduling algorithms and
Section V discusses a sensitivity study of the algorithms.
Section VI discusses the related work in this field. Section VII
discusses the conclusions and future work.
A property of MVG_OMALLOOR
ISI-TR-2006-617
10
Node 10
00
01
13101021 (Line)
M 2
11
Transmit fiber Receive fiber
Fig. 1. Typical optical CDMA LAN topology. Nodes are connected by
transmit and receive fibers to a passive optical coupler in a star topology.
II. BACKGROUND
This section provides background on optical CDMA LAN
architecture, codeset design and receiver design.
A. Optical CDMA LAN architecture
The optical CDMA network considered in this work is a
shared medium, packet switched, multiple access LAN. The
physical layer is optical CDMA that uses unipolar encoding
and intensity modulation over a single wavelength.
The network consists of several nodes connected by optical
fiber to a passive star coupler as shown in Figure 1. The optical
coupler consists of several inputs and output ports. Each node
is connected to one input and one output port by a transmit and
receiver fiber respectively. Signals transmitted on the inputs
enter the coupler, merge and are transmitted on all outputs.
The star coupler is passive i.e. the input power is split equally
among the receive fibers and is transmitted to all nodes on the
receive fibers. The signal at the output of the coupler on any
receive fiber is given by
K
X
r(t) = (1/K)
i=1
where K is the number of ports connected to the coupler and
si (t) is the signal entering on the ith transmit fiber. The signals
on the transmit fibers si (t) are binary optical signals and the
signal on a receive fiber r(t) is a multilevel optical signal.
The signal on the receive fiber may be amplified or attenuated
after the coupler.
The network is based on a Tunable Transmitter-Fixed Re-
ceiver (TT-FR) architecture. A receiver chooses a codeword to
receive on and a transmitter which needs to communicate with
a receiver tunes to the receiver’s codeword. A TT-FR architec-
ture eliminates the need for pre-transmission coordination [1].
The network uses codeword sharing. If the number of nodes is
greater than the codewords, the codewords are shared among
receivers. A higher layer unique identifier such as a link layer
address is used to demultiplex packets sharing a codeword.
Every node runs a frame synchronization algorithm [2] which
2
1 allows the node to identify that a frame destined for it has
2)
(C arrived and where the first bit of the frame begins.
B. Optical CDMA codeset design
01000011 (C1)
An Optical Orthogonal Codeset (OOC) is a set of (0,1)
sequences of length N that satisfies correlation constraints [3].
10
00
00
The term codeset is used to refer to the set of sequences, and
(C
0) the term codeword is used for a member of the set. Each 0 or
1 of a sequence is called a chip, and the codeword represents a
3 data bit. For any two codewords in the codeset, the correlation
constraints are:
N −1
X =w when i = j, τ = 0
s(i,n+τ ) s(j,n)
≤κ otherwise
n=0
where s(i,n) is the nth chip of the ith codeword, addition
is modulo N and 0 ≤ τ ≤ N − 1. κ is called the
Maximum Collision Parameter. The number w of ‘1 chips’
of a codeword of the codeset is called its weight. A particular
codeset is specified by the parameters (N, w, κ). The size S
of the codeset is the number of codewords in the codeset.
Codesets with all codewords having the same weight are called
constant weight codesets. [3] and [4] describe several codeset
construction methods. The codesets used in this work are
constant weight codesets generated by the greedy construction
method [3]. The rate at which individual chips are transmitted
is called the chipping rate B. The rate at which the data bits
are transmitted is called the data rate. The chipping rate is
N times the data rate. The codewords are pseudo-orthogonal
because optical CDMA uses unipolar encoding1.
C. Optical CDMA receiver design
The optical CDMA receiver (also called a decoder) is a
hard-limiting correlation receiver [5]. The receiver decodes the
codeword in the received signal and regenerates the transmitted
data. Figure 2 depicts the operation of a receiver. The input
signal from the coupler is a multilevel optical signal. The
receiver converts it to a digital optical signal by hard limiting
the power in each chip of the received signal. It then decodes
the signal to detect a 1 or 0 bit. Let R be the received
si (t)
signal (an N dimensional vector whose components are non-
negative integers), and C the codeword being received (an
N dimensional vector whose components are binary values).
Let R = [r0 r1 r2 ...rN −1 ] and C = [c0 c1 c2 ...cN −1 ]. Then the
received bit b is given by
b = 1 if (C · h(R) ≥ w)
= 0 otherwise
PN −1· of two vectors [u0 u1 ...uN −1 ] ·
where the dot product
[v0 v1 ...vN −1 ] = i=0 ui vi and h() is the hardlimiting
operation defined as
h(R) = [s0 s1 ...si ...sN −1 ]
where si = 0 if 0 ≤ ri < 1
si = 1 if ri ≥ 1
1 This contrasts with CDMA on the wireless medium where bipolar encod-
ing is feasible. Bipolar codewords can be designed to be orthogonal.
A property of MVG_OMALLOOR
ISI-TR-2006-617 3

0 1 1 0 0 1 0 0 C3
0 1 0 0 1 0 1 0 C2
Control
0 1 0 0 0 0 1 1 C1
0 0 0 0 0 0 0 0 C0 (0 bit)
1 1 1 1 1 3 0 1 Line
Splitter TDL Coupler
Receive
Hard- t=0 1 2 3 4 5 6 7 8 9 10
fiber Photo- Threshold
limiter TDL (a)
detector detector
0 1 1 0 0 1 0 0 C3
TDL
0 1 0 0 1 0 1 0 C2
1 0 1 0 0 0 0 1 1 C1
3 3 1
2
1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 0 0 0 1 1 0 0 C0 (1 bit)
2 1 1 1 2 4 0 1 Line

t=0 1 2 3 4 5 6 7 8 9 10
(b)

Fig. 2. Optical CDMA receiver: The figure shows a hard-limiting correlation

detector that consists of a hard-limiter, decoder, photo-detector and a threshold
detector. The receiver is tuned to the codeword 1110000. The power in the
1st , 2nd and 3rd chip positions is summed by the decoder. The photo-detector
converts the signal to an electrical signal and the threshold detector detects a
1 bit.
Fig. 3. A bit error. In (a) C0 is OFF. Codewords C1 and C2 have chip overlaps
with the 1 chips of C0. A false positive error will occur at the receiver. In
(b) C0 is ON, so no false positive error will occur.

from an (8, 3, 3) codeset2 . The figure is a snapshot of data

III. I NTERFERENCE AVOIDANCE
This section defines the transmission scheduling problem
and discusses algorithms for transmission scheduling. First
it discusses the the problem of interference, the need for
interference avoidance and how it can be implemented as
a contention media access control (MAC) protocol. Next, it
discusses state and state estimation which are needed for
transmission scheduling. Finally, it defines the transmission
scheduling problem and discusses the algorithms.
A. The need for Interference Avoidance
Interference occurs due to the multiplexing of packets on
a receive fiber. Interference errors increase as the offered
load on the network increases. Prior work [6] has shown that
without media access control, at high offered load (100%)
the throughput of the network approaches zero. The bursty
nature of data traffic means that in a operational network
there will be periods of overload. Interference Avoidance is
a distributed, contention based media access control protocol
for broadcast, packet based, shared medium optical CDMA
Local Area Networks. It improves the throughput of optical
CDMA LANs under such conditions.
When two or more packets overlap at a point on a receive
fiber (line), the codewords of the packets overlap. Codeword
overlaps may cause interference errors at the receiver. When a
codeword overlap occurs, two ‘1 chips’ of different codewords
may overlap. This is termed a chip overlap. An interference
error will occur during the reception of a codeword if there are
enough other codewords on the line which have chip overlaps
with the codeword being received. Figure 3 shows codewords
bits on an optical fiber sent by four nodes. Their combined
signal on the line is indicated below the codewords. C0 is
the codeword being received. C1 and C2 have 1 chips that
overlap with C0’s 1 chips. Figure 3(a) shows the case when a
‘0’ data bit is transmitted by the node sending C0. Figure 3(b)
shows the case when a ‘1’ data bit is transmitted. In (a)
the receiver will erroneously detect a codeword (C0) because
two other codewords overlap with it. The receiver tuned to
C0 will falsely detect a ‘1’ data bit. This results in an error
and the loss of the data packet (unless other error correction
mechanisms are used). This is an interference error. Therefore,
the condition for correct reception of a codeword is that at least
one of it’s ‘1 chips’ must not have a chip overlap with any
other codeword on the line. This works assumes that if an
interference error occurs in one bit of a packet, then the entire
packet is lost.
A simple example can illustrate the principle behind Inter-
ference Avoidance. Consider the codewords shown in Figure 4.
The codewords are from a (8, 3, 3) codeset. The signal on
the line is called the state of the line (state will be defined
formally in Section III-C). If the codeword C0 is transmitted
as shown in Figure 4(a) a false positive error would occur and
the packet sent on codeword C0 would be lost. If it was sent at
a different chip offset i.e. one chip time later (Figure 4(b)), all
three packets could be transmitted correctly. When delayed,
codeword C0 has at least one chip that does not interfere with
codewords C1, C2 and C3. Hence no false positive can occur
and it will be received correctly. Interference Avoidance uses
the above principle. A transmitting node estimates the state of
the line (state estimation) and schedules its packet transmis-
sions to avoid interference errors (transmission scheduling).
2 The figure shows the codewords as chip synchronous. In reality this may
not be true. Salehi [5] studied the effect of both chip synchronous and chip
asynchronous transmission on a correlation receiver and showed that the chip
synchronous case is a upper bound on the BER of the system. Following this
result, this work (analysis, simulation and explanations) all assume that the
codewords are chip synchronous on the fiber as shown in the figure.
A property of MVG_OMALLOOR
ISI-TR-2006-617 4

0 1 1 0 0 1 0 0 C3
0 1 0 0 1 0 1 0 C2
0 1 0 0 0 0 1 1 C1 Optical CDMA
1 0 0 0 1 1 0 0 Transmitter
C0 (1 bit) Bus
(from node Transmit
2 1 1 1 2 4 0 1 Line processor) fiber
(to coupler)
Transmission
t=0 1 2 3 4 5 6 7 8 9 10 Transmit
scheduling
buffer
module
(a) Receive
fiber
0 1 1 0 0 1 0 0 C3 (from
coupler)
0 1 0 0 1 0 1 0 C2 State State
Bus estimation observation
0 1 0 0 0 0 1 1 C1 module module
(to node
1 0 0 0 1 1 0 0 C0 (1 bit) processor)

2 1 1 1 4 1 1 1 Line
Receive Optical CDMA
buffer Receiver
t=0 1 2 3 4 5 6 7 8 9 10
(b)

Fig. 4. An interference error is caused in C0 in (a). No error is caused when Fig. 5. Block diagram of an Interference Avoidance Network Interface Card.
the same codewords are sent with a different set of chip offsets. (b) shows
codeword C0 delayed by 1 chip time

C. Optical CDMA State and State Estimation

B. Interference Avoidance media access control
Interference Avoidance is a contention media access con-
trol (MAC) protocol. Each node on the network contends
for access to the medium using the Interference Avoidance
protocol. Figure 5 shows a block digram of an Interference
Avoidance Network Interface Card. It consists of an optical
CDMA transmitter, optical CDMA receiver, state estimation
module and transmission scheduling module (state will be
formally defined in Section III-C).
The state estimation module performs two functions: receiv-
ing and estimation. It receives the multilevel optical signal on
the receive fiber and collects observations of the state of the
line. It uses a series of state observations to calculate a state
estimate. The state estimation algorithm is always active. It
is run continuously in a loop, collecting state observations
and calculating a state estimate. The state estimation module
consists of both optical and electronic components.
The transmission scheduling module uses the state estimate
and the codeword to be used for encoding to calculate a
value k (where 0 ≤ k < N ) such that interference loss
is minimized if the packet’s transmission is delayed by k
chip times relative to the packet’s arrival time. Transmission
scheduling is invoked on arrival of a packet from the node
processor. It reads the current state estimate and the codeword
for encoding and calculates the delay. The optical CDMA
transmitter encodes the data and begins transmission after the
delay. The transmission scheduling module is purely electronic
and must compute the transmission delay within a few bit
times of the current packet’s arrival.
The electronic part of the state estimation module and the
transmission scheduling module may be integrated and im-
plemented in a single ASIC chip and optimized for minimum
latency. This paper focuses on the analysis of the performance
of a network of Interference Avoidance nodes. Future work
will examine the implementation of the NIC hardware.
The signal at any time at any point on the receive fiber of
an optical CDMA LAN is a multilevel signal due to the sum
of the codewords. The state of the line is a vector of length
N equal to the sum of the codewords at output of the coupler
assuming that all nodes are transmitting 1 bits.
M
X
S(t) = [s0 s1 s2 ....sN −1 ] = rot(Ci , φi )
i=0
where M is the number of codewords on the line at the output
of the coupler at time t, Ci = [c0 c1 ...cN −1 ] is a codeword
present at the output of the coupler, rot(Ci , φi ) is a vector
of length N equal to the left rotation of the codeword Ci by
φi and φi is the number of chips between Ci ’s leading chip
(i.e. the chip that was transmitted first, c0 ) and the output of
the coupler. It is a hypothetical, idealized representation of
the state of the system. It is possible that the state may never
actually be observed as a signal on the optical fiber. The state
transitions to a new value on a packet arrival or departure.
A node can receive the multilevel optical signal and convert
it to electronic form using hardware. The received data can be
used to construct a series of state observations.
The state estimation problem is to calculate the state of the
line given a series of state observations collected from the re-
ceive fiber. A simple state estimation algorithm could estimate
the state by averaging the state observations and dividing the
result by the probability that a codeword is transmitting a 1 bit.
An appropriate higher layer encoding format (such as 8B/10B)
may be used to ensure that the observations have a sufficient
number of codewords transmitting a 1 bit.
This work assumes perfect state estimation i.e. all the nodes
on the network know the state of the line. Future work will
examine the impact of realistic state estimation.
D. Transmission scheduling
Transmission scheduling is a process by which a node, given
a state estimate and a codeword to be transmitted, calculates
a codeword delay such that interference errors are reduced. If
A property of MVG_OMALLOOR
ISI-TR-2006-617
ctx ← Codeword to be transmitted
state ← State estimate
hstate ← hardlimit(state)
td ← 0
for offset = 0 to N
if (hstate & ctx 6= ctx ) then
mark offset as a feasible offset
rotate ctx to the right by one chip
td ← any feasible offset
TABLE I
T HE P URE SELFISH TRANSMISSION SCHEDULING ALGORITHM
transmission can be scheduled, then the scheduling algorithm
returns an offset k such that 0 ≤ k < N . The offset is
the number of chips that the packet transmission should be
delayed. The offset is measured with respect to the estimated
state of the line. If transmission is not possible, then the packet
transmission is deferred by returning it to a higher layer for
a retransmission attempt. Other defer mechanisms such as
1, non and p-persistent sensing [7] may also be used. The
transmitting node does not have a receiver to detect errors in
its transmitted packet during transmission. Therefore packets
which experience interference errors during transmission are
transmitted until completion. Transmission scheduling is done
on a per packet basis.
This work assumes perfect state estimation by all nodes on
the network. In perfect state estimation, every node knows the
state of the line. All nodes see the same state at the same time.
The network is assumed to have zero propagation delay. It is
also assumed that state estimation is instantaneous and there is
no delay between state estimation and transmission scheduling.
The transmitter knows its distance from the coupler (a = 0),
therefore it can schedule its transmission using the state esti-
mate. This an idealized, unrealizable state estimation algorithm
which allows easy analysis of transmission scheduling.
When a packet is transmitted, interference errors could be
caused in itself or in other packets. One of four possible events
(transmission events) could occur:
• Preserves self/Destroys other: The packet may be re-
ceived without error (i.e. preserves itself), but may cause
an error in one or more packets (i.e. destroys others) on
the line.
• Preserves self/Preserves others: The packet preserves
itself and also preserves all other packets on the line.
• Destroys self/Preserves others: The packet destroys itself
but preserves all other packets on the line.
• Destroys self/Destroys others: The packet destroys itself
and destroys one or more other packets on the line.
When scheduling a transmission, any one of the four
transmission events could happen. A transmission scheduling
strategy is a subset of the transmission events that a node tries
to achieve. A transmission scheduling algorithm is an imple-
mentation of a strategy. Two possible transmission scheduling
strategies are selfish and cooperative. An algorithm follows a
selfish strategy if the node schedules its packet transmission
only if the packet either Preserves self/Preserves others or
Preserves self/Destroys others. An algorithm follows a coop-
5
ctx ← Codeword to be transmitted
state ← State estimate
hstate ← hardlimit(state)
td ← 0
for offset = 0 to N
if (hstate & ctx 6= ctx ) then
newstate = state + ctx
numoverlaps = overlaps(newstate)
if (numoverlaps < threshold)
mark offset as a feasible offset
rotate ctx to the right by one chip
td ← any feasible offset
TABLE II
T HE T HRESHOLD TRANSMISSION SCHEDULING ALGORITHM
erative strategy if the node schedules its transmission only if
it Preserves self/Preserves others or Destroys self/Preserves
others. Simple implementations of cooperative strategy are
either not feasible for all codesets or result in low throughput.
A pseudo-cooperative strategy attempts to reduce the proba-
bility of destroying other packets on the line. It is a best effort
strategy which increases the probability of the events Preserves
self/Preserves others or Destroys self/Preserves others.
The following sections discuss three transmission schedul-
ing algorithms: Pure selfish, Threshold and Overlap section
scheduling. The transmission scheduling algorithms imple-
ment either selfish or pseudo-cooperative strategies or both.
The section compares their performance to Aloha-CDMA i.e.
optical CDMA without any media access control.
1) Pure selfish scheduling: The pure selfish algorithm
schedules a packet transmission only if the state of the line
permits transmission without loss of its own packet. The
algorithm is specified in Table I. The algorithm searches for
chip offsets where there at least one of the ‘1 chips’ from
the codeword to be transmitted aligns with a ‘0 chip’ in the
state vector, thus ‘selfishly’ ensuring correct reception of this
codeword. It chooses one of these offsets at random.
2) Threshold scheduling: The threshold scheduling algo-
rithm searches for chip offsets where at least one of the ‘1
chips’ from the codeword to be transmitted aligns with a ‘0
chip’ in the state vector and the number of chip overlaps in
the resulting state is below a threshold. It chooses one of these
offsets at random. The threshold is expressed as a fraction of
the codeword length N , called the threshold parameter α. The
algorithm is specified in Table II.
3) Overlap section scheduling: The overlap scheduling
algorithm searches for chip offsets where at least one of the
‘1 chips’ from the codeword to be transmitted aligns with a
‘0 chip’ in the state vector and the number of chip overlaps in
the resulting state is below a threshold. It chooses one of these
offsets at random. The algorithm is specified in Table III.
IV. P ERFORMANCE STUDY
In this section the transmission scheduling algorithms are
analyzed and simulated. A mathematical analysis is described
which shows that the algorithms prevent throughput degrada-
tion. It is also shown that the threshold and overlap section
algorithms have lower packet errors compared to the pure
A property of MVG_OMALLOOR
ISI-TR-2006-617
ctx ← Codeword to be transmitted
state ← State estimate
hstate ← hardlimit(state)
td ← 0
for offset = 0 to N
if (hstate & ctx 6= ctx ) then
newstate = state + ctx
numoverlaps = overlaps(newstate)
numones = ones(newstate)
if (numoverlaps < numones)
mark offset as a feasible offset
rotate ctx to the right by one chip
td ← any feasible offset
TABLE III
T HE OVERLAP SECTION TRANSMISSION SCHEDULING ALGORITHM
selfish algorithm. A simulation study is used to validate the
mathematical analysis.
The metric used to evaluate performance is the normalized
network throughput at different values of the normalized
offered load. The normalized offered load is the arrival rate
(in packets/s) expressed as a fraction of the maximum possible
arrival rate (in packets/s) of the network when it is used as
a single channel network3. The arrival rate is defined as the
aggregate rate at which packets arrive to all the nodes for
transmission on the network. The normalized network through-
put is the ratio of the number of packets that are transmitted
over the network without error to the total number of packets
offered for transmission multiplied by the normalized offered
load. It is a measure of the throughput of packets transmitted
without error at a particular offered load. Appendix I defines
the metrics formally and derives expressions for them.
A. Analysis
This section describes a mathematical analysis of the trans-
mission scheduling algorithms. The analysis follows the steps
below:
• First, an expression for the normalized network through-
put is derived in terms of the aggregate arrival rate. This
expression is then expressed in terms of the number of
codewords on the line (i.e. at a point on the receive
fiber) and the probability of packet error (packet error
rate) (Appendix I).
• Then, a concise representation of line state which allows
easy mathematical manipulation is defined (Appendix II-
A).
• Using this state representation, expressions are derived
for the number of codewords at a point on the line and
the probability of packet error when the system is in any
state (Appendix II-B and II-C).
• Based on the transmission scheduling algorithm (Aloha-
CDMA, Pure Selfish, Threshold, Overlap section) the
state transition probabilities are calculated and a state
transition diagram is constructed. Under the assumptions
3 For an optical CDMA network of chipping rate B chip/s, the maximum
possible data rate of the network when used as a single channel network is
B b/s (the chipping rate becomes the bit rate). The maximum possible arrival
rate in packets/s is B divided by the average packet size in bits.
6
1
Threshold scheduling
Overlap section scheduling
0.9 Pure selfish scheduling
Aloha-CDMA
0.8
Normalized network throughput
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized offered load
Fig. 6. Comparison of the performance of the transmission scheduling
algorithms based on analysis. The traffic model is Poisson arrivals with expo-
nentially distributed packet lengths. The codeset is (10, 3, 3) and codewords
are chosen uniform randomly. For the threshold scheduling algorithm, the
threshold parameter was set to 0.5
of Poisson arrivals and exponentially distributed packet
sizes, the state transition diagram can be viewed as a
Markov chain. The Markov chain is solved for equi-
librium state probabilities at a particular offered load
(Appendix III).
The analysis can be used to determine the normalized
throughput at any normalized offered load. A graph of the nor-
malized throughput vs. normalized offered load for different
scheduling algorithms is shown in Figure 6. The traffic model
is Poisson arrivals and exponentially distributed packet sizes.
Codeword allocation is uniform random over the codeset. The
graph indicates that a system with no transmission scheduling
(Aloha-CDMA) suffers throughput degradation. This may be
seen from the performance at high offered loads. Beyond
an offered load of around 0.5, the network throughput de-
creases and tends to zero at high offered loads. In contrast,
the transmission scheduling algorithms all prevent throughput
degradation. Throughput is stabilized at around 30% of the
maximum throughput and remains stable as offered load is
increased.
B. Simulation
A discrete event based packet simulator was designed to
validate the mathematical analysis. The simulator modeled
multiple nodes on a broadcast shared medium optical CDMA
LAN. It implemented different state estimation algorithms,
transmission scheduling algorithms and a hard-limiting cor-
relation receiver. Unless specified otherwise, the default pa-
rameters for the simulations are specified in Table IV. The
optical orthogonal codeset construction method was the greedy
construction method [3]. It was used to generate several
codesets for a given set of codeset parameters. The results
in this work did not depend on the codeset or the algorithm
used to generate the specific optical orthogonal codeset. The
results from the simulations are the mean of around 10 runs of
anywhere from 10,000 to 100,000 packets each and standard
deviations are shown on the graphs.
A property of MVG_OMALLOOR
ISI-TR-2006-617 7

1 10
Threshold scheduling Threshold scheduling
Overlap section scheduling Overlap section scheduling
0.9 Pure selfish scheduling 9 Pure selfish scheduling
Aloha-CDMA Aloha-CDMA
0.8 8

multiplexed at a point on the line

Normalized network throughput

Average number of codewords

0.7 7

0.6 6

0.5 5

0.4 4

0.3 3

0.2 2

0.1 1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized offered load Normalized offered load

Fig. 7. Comparison of the performance of the transmission scheduling
algorithms based on simulation. The traffic model is Poisson arrivals with
exponentially distributed packet lengths. The codeset is (10, 3, 3) and code-
words are chosen uniform randomly. For the threshold scheduling algorithm,
the threshold parameter was set to 0.5
Fig. 8. Comparison of the codeword multiplexing of the different transmis-
sion scheduling algorithms based on simulation. The traffic model is Poisson
arrivals with exponentially distributed packet lengths. The codeset is (10, 3, 3)
and codewords are chosen uniform randomly. For the threshold scheduling
algorithm, the threshold parameter was set to 0.5

1
Threshold scheduling
Figure 7 shows the results of simulation for the same 0.9
Overlap section scheduling
Pure selfish scheduling
codeset as described in the analytical results. The results Aloha-CDMA
0.8
are quite similar to the analytical results. All transmission
0.7
scheduling algorithms prevent throughput degradation. Also Packet error rate

the overlap section and threshold scheduling show marginally 0.6

higher throughput than pure selfish scheduling. The analytical 0.5

model over predicts the throughput for Aloha-CDMA. This is 0.4

because the analysis is based on a finite state model. A finite 0.3

state model is suitable for transmission scheduling algorithms 0.2
which limit the traffic on the line. However for Aloha-CDMA, 0.1
the finite state model over predicts the equilibrium state
0
probabilities. As a result the analytical results differ from 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized offered load
simulation.
Though Figures 6 and 7 indicate that all three scheduling Fig. 9. Comparison of the packet error rate for different transmission
scheduling algorithms based on simulation. The traffic model is Poisson
algorithms has approximately the same throughput, the algo- arrivals with exponentially distributed packet lengths. The codeset is (10, 3, 3)
rithms differ in the packet error rate. Figures 8 and 9 show and codewords are chosen uniform randomly. For the threshold scheduling
the average number of codewords multiplexed on the line at a algorithm, the threshold parameter was set to 0.5
point on a receive fiber and the average packet error rate for
the transmission scheduling algorithms.
There is a trade-off between the number of codewords at a number of codewords on the line. To the right of the maximum,
point on the line and the packet error rate. The transmission the throughput is lower due to the higher packet error rate. The
scheduling algorithm tells a node if it can transmit. It also objective of the transmission scheduling algorithm is to keep
tells the node when to transmit. By doing this the transmission the system operating point close to the maximum irrespective
scheduling algorithm controls two quantities: of the offered load. Therefore, if the transmission scheduling
• The number of codewords at a point on the line. algorithm is aggressive and allows more codewords on the
• The chip offsets of the codewords on the line which line, the packet error rate increases, lowering throughput. On
affects the packet error rate. the other hand if the algorithm is conservative and does not
The trade-off can be understood by considering the perfor- allow enough codewords on the line, the throughput remains
mance of Aloha-CDMA. As the offered load increases, the low. So the transmission scheduling algorithm must carefully
number of codewords on the line for Aloha-CDMA shows balance the number of codewords on the line and the codeword
a linear increase with offered load. Initially as the number offsets so as to maximize throughput.
of codewords on the line increases, the network throughput The difference between the three scheduling algorithms is
increases. But as the number of codewords on the line in- evident from Figure 9. There are differences in the packet error
creases further, interference errors increase and as a result the rate. The threshold and overlap section scheduling algorithms
packet error rate increases. As a result the network throughput are conservative and constrain the number of overlapping
falls. Therefore as the offered load increases, the network chips. This results in a lower number of codewords on the
throughput attains a maximum and then decreases. To the line and low errors due to interference. However the self-
left of the maximum, the throughput is lower due to the low ish algorithm is aggressive and admits a larger number of
A property of MVG_OMALLOOR
ISI-TR-2006-617 8

Parameter Default value 1

Codeset parameters: Threshold scheduling
Overlap section scheduling
0.9
Codeset length N 100 Pure selfish scheduling
Aloha-CDMA
Number of wavelengths Λ 1 0.8

Normalized network throughput at

Codeset weight w 3

normalized offered load of 1

0.7
Maximum crosscorrelation parameter κ 3
Size of codeset S 100 0.6
Chipping rate: 10 Gc/s
0.5
Codeword allocation: Uniform random
Interference Avoidance parameters: 0.4
Transmission scheduling algorithm: Threshold scheduling
0.3
Threshold: 0.5
State estimation algorithm: Perfect state estimation 0.2
Traffic parameters:
0.1
Inter-arrival time distribution Exponential
Normalized offered load 1 0
0 10 20 30 40 50 60 70 80 90 100
Packet size distribution Exponential
Codeset length N
Average packet size 1000 bytes
Destination address distribution: Uniform random Fig. 10. Comparison of the performance of the transmission scheduling
Topology parameters: algorithms as the codeset length is varied (based on simulation). The traffic
Node to coupler distance distribution Deterministic model is Poisson arrivals with exponentially distributed packet lengths. The
Average node to coupler distance 0m codeset weight is 3 and κ = 3. Codewords are chosen uniform randomly
Number of nodes 100 from the set. For the threshold scheduling algorithm, the threshold parameter
was set to 0.5
TABLE IV
PARAMETER LIST AND DEFAULT VALUES FOR THE TRANSMISSION
SCHEDULING SENSITIVITY STUDY. codeset length. Therefore, network throughput is independent
of codeset length and transmission scheduling scales with
increase in codeset length.

codewords with a higher probability of packet error. When
packets are lost due to interference errors, the higher layers
of the protocol stack must recover through some form of
Automatic Repeat Request (ARQ) or Forward Error Correction
(FEC). Therefore the threshold and overlap section scheduling
algorithms are better choices because a larger fraction of the
packets transmitted are transmitted without error.
V. S ENSITIVITY ANALYSIS
A simulation based study was conducted to allow a deeper
sensitivity analysis of the transmission scheduling algorithms.
The study considered parameters at the physical layer (codeset
parameters), the media access control layers (the scheduling
algorithm parameters) and the traffic model (packet arrival and
size distributions). The objective of this study is to quantify
the impact of these factors on the transmission scheduling
algorithms. The sensitivity analysis consisted of quantifying
the:
• Effect of varying the codeset length.
• Effect of varying the codeset weight.
• Effect of different packet size distributions.
• Performance under realistic network traffic.
B. Effect of varying the weight of the codeset
Changing the codeset weight w has two effects: As w
is increased, the threshold on the correlation receiver can
be increased. Increasing the threshold makes the codewords
more resistant to interference. This is because it takes a
larger number of chip overlaps to cause a false positive error.
However, with the increase in weight, each codeword causes
more interference with other codewords. An increase in w also
makes it difficult for the transmission scheduling algorithm to
schedule codewords on the line without causing interference
errors. This reduces the number of codewords that can be
simultaneously transmitted on the line. The results show that
as the weight increases the throughput decreases rapidly. The
reduction in the codewords on the line and the increased
interference offset any gains in the resistance to interference.
It has been shown [8] that codesets with higher weight have
better bit error rate characteristics at low loads. At low loads
the increase in resistance to interference dominates resulting
in higher throughput for high weight codesets. However as the
load is increased, the effect of interference tends to dominate,
resulting in lower throughput for high weight codesets. Fig-
ure 11 shows a graph of the network throughput at offered
load of 1 vs. the weight.

A. Effect of varying the length of the codeset
Increasing the length of the codeset has several effects. As
N increases, the scheduling algorithm can schedule a larger
number of codewords simultaneously on the line. Therefore
more nodes can transmit in parallel without error. However
the nodes transmit at a lower data rate. The results show
that the two effects balance each other and the network
throughput is constant when N is varied. Figure 10 shows
a graph of network throughput at an offered load of 1 vs. the
C. Effect of packet size distribution
Figure 12 shows the packet throughput as the average
packet size is varied. The traffic model is Poisson arrivals
and exponentially distributed packet sizes. The figure shows
that the packet size has no effect on packet throughput.
However, studies indicate that real network traffic packet
size distributions may not be exponential [9]. Recent packet
statistics obtained from a backbone network [10] exhibit a
trimodal distribution. In one trace, about 70% of the packet
A property of MVG_OMALLOOR
ISI-TR-2006-617 9

1 1
Threshold scheduling Threshold scheduling
Overlap section scheduling Overlap section scheduling
0.9 Pure selfish scheduling 0.9 Pure selfish scheduling
Aloha-CDMA Aloha-CDMA
0.8 0.8
Normalized network throughput at

Normalized network throughput

normalized offered load of 1

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 1 2 3 4 5 6 7 8 9 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Codeset weight w Normalized offered load

Fig. 11. Comparison of the performance of the transmission scheduling
algorithms as the codeset weight is varied (based on simulation). The traffic
model is Poisson arrivals with exponentially distributed packet lengths. The
codeset length is 100 and κ = 3. Codewords are chosen uniform randomly for
the codeset. For the threshold scheduling algorithm, the threshold parameter
was set to 0.5
Fig. 13. Comparison of the performance of the transmission scheduling
algorithms for a trimodal packet size distribution (based on simulation). The
traffic model is Poisson arrivals with packet size distribution consisting of
70% 40 byte packets, 20% 1500 byte packets and 10% 500 byte packets. The
codeset is (100, 3, 3) and codewords are chosen uniform randomly. For the
threshold scheduling algorithm, the threshold parameter was set to 0.5

1
Threshold scheduling
Overlap section scheduling
where short packets tend to experience lower error rates than
0.9 Pure selfish scheduling
Aloha-CDMA long packets [11].
0.8
The squeeze through effect can be demonstrated analytically
Normalized network throughput at
normalized offered load of 1

0.7 and through simulation for a network with a bimodal distri-

0.6
0.5
0.4
0.3
0.2
0.1
0
0 100 200 300 400 500 600
Mean packet size in bytes
Fig. 12. Comparison of the performance of the transmission scheduling
algorithms as the average packet length is varied (based on simulation). The
traffic model is Poisson arrivals and exponentially distributed packet sizes.
The codeset is (100, 3, 3) and codewords are chosen uniform randomly. The
algorithm was threshold scheduling, the threshold parameter was set to 0.5
sizes were 40 bytes, about 20% were 1500 bytes and the
remaining were around 500 bytes long. A traffic model with
such a trimodal packet size distribution was used to drive a
simulation that used Poisson arrivals, uniform random codeset
allocation and a (100, 3, 3) codeset. The results are shown
in Figure 13. The graph shows interesting behavior. Aloha-
CDMA does not degrade as much as in the case of Pois-
son traffic/exponential packet sizes. The other transmission
scheduling algorithms have almost 25% higher throughput
when compared to performance with exponentially distributed
packet sizes. A majority of the packets are small size packets
(40 bytes). Study indicates that the packet error rate is lower
for short packets than for long packets. Fewer long packets
are transmitted on the line and a large fraction of them are
lost due to errors. This squeeze through effect results in an
increase in the aggregate network packet throughput. This
behavior occurs when the fraction of shorter packets is fairly
high (60-70%). This is similar to behavior on wireless links
bution of packet sizes which uses a pure selfish transmission
scheduling algorithm and a codeset with κ = w. Consider a
network where the traffic has two packet types of sizes l1
and l2 where l1 < l2 . Let the fraction of packets of size
l1 be γ. Then the average packet size on the network is
lavg = (γ)l1 + (1 − γ)l2 . The throughput of such a network
can be calculated by finding the probability of packet error
700 800 900 1000 Perror and the number of codewords on the line Nl .
The average number of codewords on the line at any point
on the receive fiber Nl depends on the transmission scheduling
algorithm. If there are Nl codewords on the line, then Nl w
1 chips were added to the state. Of these, Nl + (w − 1) 1
chips were aligned with 0 chips of the state (during selfish
scheduling) and (Nl −1)(w −1) 1 chips were added randomly
to any position (κ = w). The probability that the state vector
has no 0 chips is
1
Pf ull = 1 − (1 − )(Nl −1)(w−1)
(N − (Nl + (w − 1)))
At an offered load of 1, packets are arriving for transmission
at a rate much higher than the rate at which packets are trans-
mitted (the transmission scheduling does not allow all packets
to be transmitted). When a packet departs from the line, a few
chips of the state may change from 1 to 0. The next packet
arrival will result in a transmission of a packet such that the
0 chips will be filled. Therefore under equilibrium conditions,
for the pure selfish scheduling algorithm, Pf ull will be close
to 1. Through simulation Pf ull is determined to be around
0.85 for the pure selfish algorithm under Poisson arrivals and
exponentially distributed packet sizes. It is assumed that this
is true for bimodal packet distributions too. Therefore, the
number of codewords on the line can be calculated by finding
Nl such that Pf ull is close to 1.
Interference errors in a packet on the line are caused
A property of MVG_OMALLOOR
ISI-TR-2006-617
1 1
0.9
0.8
Normalized network throughput
0.7
at an offered load of 1
0.6
0.5
0.4
0.3
0.2
0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6
Fraction of short packets
Fig. 14. The squeeze through effect. The throughput is maximized when
the fraction of short packets is 0.9. The graph shows both analytical and
simulation results. The transmission scheduling algorithm is pure selfish. The
packet sizes are 50 bytes and 1000 bytes. The codeset is (100,3,3).
by packets that arrive during the packet’s transmission. The
transmission scheduling algorithm allows only a fraction of the
arriving packets (called colliding packets) to be transmitted.
The probability of packet error in a codeword on the line is
the probability that at least one of its colliding packets causes
an interference error. If the number of colliding packets is nc ,
then nc 1 chips are added to the state selfishly (align with 0
chips) and nc (w − 1) 1 chips are added to the state in random
positions. These random positions are chosen from N − 1
possible choices (1 chip is chosen selfishly). The probability
that one of the added 1 chips overlaps with the one of the 1
chips of the codeword on the line is p = w/(N − 1). The
probability of packet error is the probability that more than
w − 1 overlaps occur. Therefore,
w−1
X  nc (w − 1) 
Perror = 1 − pk (1 − p)nc (w−1)−k
k
k=0
At a normalized offered load of 1, the average packet inter-
arrival time is tarrival = lavg /B where B is the chipping rate
of the network. The transmission time for packets of type 1 is
t1 = l1 /(B/N ) where N is the codeset length. The average
number of colliding packets for packet type 1 is,
nc1 = (t1 /tarrival ) ∗ (Nl /N )
A similar expression can be derived for nc2 . This can be used
to calculate the probability of packet error for each type of
packet Perror1 and Perror2 .
The normalized network throughput based on the definition
in Appendix I is given by
T h = (Nl /N )(γ(1 − Perror1 ) + (1 − γ)(1 − Perror2 ))
A graph of normalized network throughput vs. the fraction
of short packets is shown in Figure 14. The packet sizes were
set to 50 bytes and 1000 bytes. The results (both analysis
and simulation) show that the normalized network throughput
peaks at a particular value of the fraction of small packets
confirming the squeeze through effect. The long packets expe-
rience high packet error rates and the short packets experience
10
Simulation Threshold scheduling
Analysis Overlap section scheduling
0.9 Pure selfish scheduling
Aloha-CDMA
0.8
Normalized network throughput
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized offered load
Fig. 15. Comparison of the performance of the transmission scheduling
algorithms for a realistic traffic model (based on simulation). The traffic
model was based on real network traffic traces (see description). The codeset
is (100, 3, 3) and codewords are allocated to addresses. For the threshold
scheduling algorithm, the threshold parameter was set to 0.5
low error rates. The throughput attains a maximum when the
fraction of short packets reaches a particular value (around
0.9).
The higher throughput of shorter packets may not be a
desirable characteristic, because it is unfair to longer packets.
Future work will address the issue of providing a uniform
dropping probability to all traffic. Possible alternatives include
using constant packet sizes or varying the codeset length.
D. Performance with real network traffic
The assumption of Poisson arrival and exponential dis-
tributed packets lengths is convenient for analysis. However,
the inter-arrival and packet size distribution of real network
traffic could be different depending on when and where it is
observed.
Simulations were performed with traffic traces obtained
from a real network link to understand the impact of real
packet arrival times. Traffic traces from a single OC48 [12]
link were used. Several of these traces were merged to generate
traffic of different offered loads. The packet sizes, source
addresses, and destination addresses were preserved during
merging. The packet size distribution was trimodal (35% of
the packets were 40 bytes, 30% were 500 bytes and 35% were
1500 bytes). The traffic file had approximately 6000 unique
source addresses and 40000 unique destination addresses. In
contrast to all the previously described results, in this case
the nodes mapped destination addresses to codewords before
transmission. Where codewords were insufficient, codeword
reuse was used.
The results of the simulation are shown in Figure 15. The
results indicate that the performance is similar to that of
the Poisson arrivals/exponentially distributed model, indicating
that it was a fairly reasonable choice for analysis. Note that
the performance improvements of the previous section due to
the squeeze through effect are not visible here. The proportion
of smaller packets (40 bytes) in the traffic was not sufficient
to cause the squeeze through effect.
A property of MVG_OMALLOOR
ISI-TR-2006-617
VI. R ELATED WORK
Work related to Interference Avoidance can be divided into
four areas: Bit error rate analysis of optical CDMA networks,
optical CDMA codeset design, FEC for optical CDMA and
media access control in optical/wireless networks.
Salehi [5], [8] analyzed an optical CDMA based network
and developed expressions for the bit error rate of a network
that uses codesets with κ = 1. The analysis also determined
the bit error rate for codesets with different lengths and
weights and with hard-limiting at low loads using Aloha-
CDMA. This work examines these results in the context of
transmission scheduling at high offered loads.
The area of optical CDMA code design has focused on con-
struction of codesets with large size. Chung et al. [3] described
several algorithms to construct OOCs. These constructions are
for codes with maximum crosscorrelation parameter κ = 1.
Chung and Kumar [13] described a method for construction
of codes with κ = 2. Several construction methods for OOCs
are described in [4] and [14] among others.
Efforts at reducing packet errors in optical CDMA have
mostly focused on using error correcting codes on top of
optical CDMA. Hsu et al. [15] analyzed the performance of
slotted and unslotted optical CDMA packet networks. They
developed expressions for the throughput of the network and
showed performance can be improved using Forward Error
Correction (FEC) codes and hard limiters. Muckenheim et
al. [16] studied the effect of bit error probability on the packet
error probability and suggested the use of block codes to
reduce packet errors. They also described a random delay
protocol to reduce the errors incurred during periods of high
activity and showed throughput improvement. The mechanism
detects periods of high activity and defers transmissions. In
contrast Interference Avoidance does not use any FEC and
schedules packet transmissions to avoid interference.
Contention media access control mechanisms such as Car-
rier Sense and Collision Detection for optical networks have
been studied in [17]. Interference Avoidance in optical CDMA
is analogous to carrier sensing in single channel shared
medium networks. Interference Avoidance can be viewed as
the result of applying CSMA design principles to the optical
CDMA physical layer.
Synchronous-CDMA [18] is a CDMA based physical layer
for cable networks. Nodes reserve codewords to use in a time
slot through a centralized controller. This mechanism is similar
to the Interference Avoidance techniques in that it performs
admission control and restricts the number of simultaneous
users on the network. However it is centralized and admission
control is static (by codeword assignment).
In wireless networks, several mechanisms have been pro-
posed to take advantage of knowledge about channel condi-
tions. State in wireless networks has to account for the chan-
nel noise and multi-path effects. Channel load sensing [19],
[20], [21], MIMO/CSIT based media access control [22] and
scheduling algorithms [23] are systems where wireless state
has been used to schedule transmissions. The differences
between these mechanisms and Interference Avoidance are
the state description and the scheduling algorithms. In these
11
systems state is a scalar variable and media access control
is through admission control i.e. the number of simultaneous
users is controlled.
VII. C ONCLUSIONS AND FUTURE WORK
This work has presented an analysis of transmission
scheduling algorithms for optical CDMA media access control.
The analysis quantified the difference between throughput of
systems with and without transmission scheduling and showed
that transmission scheduling achieved 30% throughput while
non scheduled systems had close to zero throughput. Simula-
tions showed that the throughput of transmission scheduling is
independent of codeset length. It also showed that an increase
in weight can lead to a degradation in the performance of these
algorithms, although the degradation is not as bad as systems
without transmission scheduling. Simulations also showed that
transmission scheduling prevents degradation when used with
a realistic traffic model based on traffic obtained from a real
network.
Limitations of this work include the fact that it assumes per-
fect state estimation and neglects errors due to synchronization
and receiver contention. Future work will explore the impact
of realistic state estimation.
Work in progress includes a testbed implementation of the
transmission scheduling hardware. The testbed demonstrates
a simplified form of threshold transmission scheduling by
transmitting bits such that the number of chip overlaps is
constrained. Measurements indicate that the bit error rate is
substantially lower for this system than for a system without
transmission scheduling.
R EFERENCES
[1] B. Mukherjee, “WDM based local lightwave networks. Part I: Single
hop systems,” IEEE Network Magazine, vol. 6, no. 3, pp. 12–27, May
1992.
[2] A. Keshavarizian and J. A. Salehi, “Optical orthogonal code acquisi-
tion in fiber optic CDMA systems via simple search method,” IEEE
Transactions on Communications, vol. 50, no. 3, pp. 473–483, March
2002.
[3] H. Chung, J. Salehi, and V. K. Wei, “Optical orthogonal codes: Design,
analysis, and applications,” IEEE Transactions on Information theory,
vol. 35, no. 3, pp. 595–605, May 1989.
[4] O. Moreno, Z. Zhang, P. Kumar, and V. Zinoviev, “New constructions
of optimal cyclically permutable constant weight codes,” IEEE Trans-
actions on Information Theory, vol. 41, no. 2, March 1995.
[5] J. Salehi, “Code division multiple-access techniques in optical fiber
networks - Part 1: Fundamental principles,” IEEE Transactions on
Communications, vol. 37, no. 8, pp. 824–833, Aug. 1989.
[6] P. Kamath, J. D. Touch, and J. A. Bannister, “The need for media access
control in optical CDMA networks,” in IEEE Infocom, vol. 4, March
2004, pp. 2208–2219.
[7] F. A. Tobagi and V. B. Hunt, “Performance analysis of carrier sense
multiple access with collision detection,” Computer Networks, vol. 4,
no. 5, pp. 245–259, Oct. 1980.
[8] J. Salehi and C. Brackett, “Code division multiple-access techniques in
optical fiber networks - Part 2: Systems performance analysis,” IEEE
Transactions on Communications, vol. 37, no. 8, pp. 834–842, Aug.
1989.
[9] V. Paxson and S. Floyd, “Wide area traffic: the failure of Poisson
modeling,” IEEE/ACM Transactions on Networking, vol. 3, no. 3, pp.
226–244, 1995.
[10] Sprint ATL Labs, “Packet size distributions from an internet backbone,”
www.sprintlabs.com, March 2005.
A property of MVG_OMALLOOR
ISI-TR-2006-617
[11] B. S. Bakshi, P. Krishna, N. H. Vaidya, and D. K. Pradhan, “Improving
performance of TCP over wireless networks,” in International Confer-
ence on Distributed Computing Systems (IDCS), May 1997, pp. 365–
373.
[12] NLANR, “Traffic traces from an OC48 link,” www.nlanr.net, March
2005.
[13] H. Chung and P. V. Kumar, “Optical orthogonal codes - new bounds
and an optimal construction,” IEEE Transactions on Information Theory,
vol. 36, no. 4, pp. 866–873, July 1990.
[14] R. Fuji-Hara and Y. Miao, “Optical orthogonal codes: Their bounds and
new optimal constructions,” IEEE Transactions on Information Theory,
vol. 46, no. 7, pp. 2396–2406, Nov. 2000.
[15] C. S. Hsu and V. O. K. Li, “Performance analysis of slotted fiber-
optic code division multiple access (CDMA) packet networks,” IEEE
Transactions on Communications, vol. 45, no. 7, pp. 819 – 828, July
1997.
[16] J. Muckenheim, K. Iversen, and D. Hampicke, “Construction of high-
efficient optical CDMA computer networks: Statistical design,” in IEEE
International Conference on Communications, vol. 3, June 1998, pp.
1289–1293.
[17] C. Chae, E. Wong, and R. Tuckker, “Ethernet over passive optical
network based on optical CSMA/CD media access technique,” in In-
ternational Symposium on Services and Local access, April 2002.
[18] Cable Television Labs Inc., “Data Over Cable Service Interface Spec-
ifications (DOCSIS) 2.0, Radio Frequency Interface Specification,” no.
SP-RFIv2.0, April 2004.
[19] A. H. Abdelmonem and T. N. Saadawi, “Performance analysis of spread
spectrum packet radio network with channel load sensing,” IEEE Journal
on Special Areas in Communications, vol. 7, no. 1, pp. 161–166, Jan.
1989.
[20] G. Judge and F. Takawira, “Spread Spectrum CDMA Packet Radio MAC
Protocol using Channel Overload Detection and Blocking,” Wireless
Networks, vol. 6, pp. 467–479, December 2000.
[21] K. Toshimitsu, T. Yamazoto, M. Katayama, and A. Ogawa, “A novel
spread slotted Aloha system with channel load sensing protocol,” IEEE
Journal on Special Areas in Communications, vol. 12, no. 4, pp. 665–
672, May 1994.
[22] S. A. Reddy and L. Tong, “Exploiting decentralized channel state infor-
mation for random access,” IEEE Transations on Information Theory,
vol. 51, no. 2, pp. 537–561, Feb. 2005.
[23] T. Shepard, “Decentralized channel management in scalable multihop
spread-spectrum packet radio networks,” PhD. Thesis, no. MIT/LCS/TR-
670, 1995.
A PPENDIX I
O PTICAL CDMA LAN PERFORMANCE METRICS
Let the aggregate arrival rate of packets to the network be
λ packets/s. If the average packet size is L bytes, the chipping
rate is B chips/s, then the maximum possible packet arrival
rate on a single channel network is B/8L packets/s. Therefore,
the normalized offered load ρ is,
ρ = 8Lλ/B
Consider an optical CDMA LAN where the arrival rate of
packets is λ packets/s. The service rate µ (in packets/s) of the
network is the rate at which packets are transmitted on the
network. For an optical CDMA LAN using a codeset (N ,
w, κ), an average packet size of L bytes, and a chipping
rate B chips/s the service rate is µ = B/(8LN ) packets/s.
Substituting for B in the expression above,
ρ = 8Lλ/(8LN µ)) = λ/µN
Consider an Aloha-CDMA system where packets are trans-
mitted on arrival. Consider any point on a receive fiber. If each
packet takes an average time of 1/µ seconds at that point, then
Little’s law, the average number of packets at that point (on
the line) is λ/µ.
12
For any other transmission scheduling algorithm, the same
average number of packets are offered for transmission. How-
ever the transmission scheduling algorithm does not allow all
of these packets to be transmitted. Let the average number
of packets on the line at any point of a receive fiber be
Nonline , of which a fraction Pe are lost due to error. The
ratio of the average number of error free packets transmitted
to the average number of packets offered for transmission is
Nonline (1 − Pe )/(λ/µ)
Therefore,
T hnorm = (Nonline (1 − Pe )/(λ/µ))ρ
= (Nonline (1 − Pe )/(λ/µ))(λ/µN )
= Nonline (1 − Pe )/N
A PPENDIX II
N ORMALIZED NETWORK THROUGHPUT
This appendix derives an expression for the normalized
network throughput of an Interference Avoidance based optical
CDMA LAN. First, a concise representation of line state which
allows easy mathematical manipulation is defined. Using this
state representation, expressions are derived for the number of
codewords on the line and the probability of error when the
system is in any state.
A. State representation
The state of the line can be represented by a pair (n0 , n1 )
where n0 < N and n1 < N . n0 is the number of zeros
in the true state and n1 is the number of ones. The number
of overlaps is nov = N - (n1 + n0 ). The term state will be
used to refer to this reduced representation of the state of the
line. The state of the line could be any value (n0 , n1 ) where
0 ≤ n0 ≤ N and 0 ≤ n1 ≤ N . A valid state is defined
as a state where n0 + n1 ≤ N . All other states are invalid
(when n0 + n1 > N ) The initial state is defined as the state
of the line when no codewords are on the line i.e. (N, 0). A
reachable state is defined as a state which can be reached from
the initial state by a series of state transitions due to packet
arrivals. The initial state (N, 0) is, by definition, a reachable
state. The set of reachable states depends on the transmission
scheduling algorithm. When there are no codewords on the
line i.e. no node is transmitting packets, the state at a point
on the line is the initial state i.e. (N , 0), i.e. N zeroes and
0 ones and overlaps. The arrival of a single codeword adds
w ‘1 chips’ to the state of the line and the state changes.
This is called a state transition. Let the start state of a state
transition be (f romn0 , f romn1 ) and the destination state be
(ton0 , ton1 ). A state transition may be caused only by a packet
arrival or departure4. When a codeword is added to the line,
w 1 chips are added. A ‘1 chip’ could overlap with a 0, 1 or
an overlap. Let the number of ‘1 chips’ overlapping with 0s,
1s and overlaps be c0 , c1 and coverlap respectively.
4 The effect of ON-OFF keyed modulation is neglected by assuming that a
packet consists of only 1 data bits i.e. all data bits are ON. Packet arrivals
and departures are assumed to be the sole cause of any state change. This
assumption means that the probability of error calculated is higher that the
true value and the throughput is the worst case throughput.
A property of MVG_OMALLOOR
ISI-TR-2006-617 13

assumed to be ON i.e. transmitting 1 bits although that may

N-w-1, w N-w, w
not necessarily be true. Therefore the calculated packet error
rate is the worst case packet error rate.
N-w-1, w-1 N-w-1, w-1

C. Number of codewords multiplexed at a point on the line

N-w-2,w-2
N, 0
Fig. 16. State transition diagram
Then,
c0 = f romn0 − ton0
c1 = tonoverlap − f romnoverlap
coverlap = w − (c0 + c1 )
A valid transition is defined as a transition where the start
state and the destination state are valid, reachable states and
c0 ≥ 0
c1 ≥ 0
coverlap ≥ 0
An admissible transition is defined as a valid transition
which is permitted by the transmission scheduling algorithm.
A same state transition is defined as a transition from a state
to itself. A state transition diagram can be drawn based on
the admissible transitions. Figure 16 shows a portion of a state
transition diagram for a (N, w, κ) codeset. Because of the large
size of the diagram, only a few states are shown. Invalid states
are not indicated in the diagram. The diagram shows the initial
state (N, 0) and state transitions to a few states.
Consider a graph where the nodes of the graph are the states
and each state transition due to an arrival forms a directed edge
(neglect same state transitions). For any selfish transmission
scheduling algorithm, this graph is a directed acyclic graph.
In the graph, each edge represents the arrival of exactly one
packet. Therefore the number of codewords on the line for
a state (n0 , n1 ) depends on the number of edges from the
initial state (N, 0) to (n0 , n1 ). Both the shortest and the longest
path may be calculated in polynomial time. The length of the
shortest and longest path from the initial state to that state are
lower and upper bounds on the number of codewords on the
line when the line is in that state5 .
Therefore, for valid states,
Nonline (n0 , n1 ) ≥ ShortestP ath((N, 0), (n0 , n1 ))
Nonline (n0 , n1 ) ≤ LongestP ath((N, 0), (n0 , n1 ))
For invalid states,
Nonline (n0 , n1 ) = 0
D. Normalized network throughput
From Appendix I, the normalized throughput in a state (n0 ,
n1 ) is given by

T h(n0 , n1 ) = (Nonline (n0 , n1 )/N )(1 − Pe (n0 , n1 ))

B. Probability of packet error
The probability of error depends on the state. For a state
(n0 , n1 ), the probability of error is the probability that for
any codeword, all its ‘1 chips’ overlap with other 1 chips or
overlaps. To calculate the probability of error, the locations of
the ‘1 chips’ in the codeword are assumed to be chosen random
i.e. the codeset construction uses κ = w. For any selfish
algorithm, in state (n0 , n1 ), if Nonline (n0 , n1 ) codewords are
multiplexed at a point on the line, then at most n1 codewords
are transmitted without error. Therefore the probability of
packet error can be approximated by the expression below:
For valid states,
Pe (n0 , n1 ) =0 when Nonline (n0 , n1 ) ≤ n1
(Nonline (n0 , n1 ) − n1 )
= otherwise
Nonline (n0 , n1 )
For invalid states,
Pe (n0 , n1 ) = 0
Note that a packet is considered lost if there are other
codewords on the line whose chips overlap with all w chips
of the packet’s codeword. The other codewords on the line are
where Nonline (n0 , n1 ) is the number of codewords on the
line at a point on the receive fiber in any state and Pe (n0 , n1 )
is the probability of error in any state.
The average normalized throughput at a given offered load
is given by
N N
X X
T hnorm = Pstate (n0 , n1 )T h(n0 , n1 )
n0 =0 n1 =0
where Pstate (n0 , n1 ) is the probability of being in state
(n0 , n1 ) at equilibrium.
A PPENDIX III
E QUILIBRIUM STATE PROBABILITIES
The equilibrium state probabilities are calculated by mod-
eling state transitions as a Markov chain.
5 This work assumes the number of codewords on the line is equal to
the lower bound. Therefore it is a worst case assumption and the calculated
throughput is a lower bound on the achievable throughput.
A property of MVG_OMALLOOR
ISI-TR-2006-617 14

A. Assumptions E. Balance equations

The analysis assumes perfect state estimation. The only When the system is in equilibrium, the flow into any state
reason for a state transition is an arrival or a departure of a must equal the flow out of the state. Therefore for a valid,
packet. Packet arrivals are assumed to be Poisson arrivals and reachable state (s0 , s1 ),
packet lengths are exponentially distributed. The distribution
(λ + N (s , s )µ)Pstate (s0 , s1 ) =
of the destination’s codeword is uniform over the codeset. PN 0 1
PN online
Under this assumption, the probability of transitioning to a PNn0 =0 Pn1 =0 Pstate (n0 , n1 )λParr (n0 , n1 , s0 , s1 )+
N
particular state on an arrival is dependent only on the current n0 =0 n1 =0 Pstate (n0 , n1 )Nonline (n0 , n1 )µPdep (n0 , n1 , s0 , s1 )
state and not on the path taken to reach that state. The Also,
probability of departure to a state is assumed to be proportional N
X N
X
to the rate of arrival from that state. Then the next state Pstate (n0 , n1 ) = 1;
is dependent only on the current state and not on the path n0 =0 n1 =0

taken to reach that state. Under these circumstances, the state These equations can be solved for the equilibrium state
transition diagram for arrivals and departures is a Markov probabilities Pstate (s0 , s1 ).
chain. Equilibrium probabilities may be found by solving the
balance equations for the system.

B. Admissible transmissions
This subsection describes how to identify admissible state
transitions given a codeset and a transmission scheduling
algorithm.6
1) Aloha-CDMA: For Aloha-CDMA, all transitions are
admissible.
2) Pure selfish scheduling: A transmission is admissible if,
c0 ≥ 1
3) Threshold scheduling: A transmission is admissible if
c0 ≥ 1
tonoverlap ≤ αN
where α is the threshold parameter.
4) Overlap section scheduling: A transmission is admissi-
ble if
c0 ≥ 1
tonoverlap ≤ ton1

C. Arrival state transition probabilities

For an admissible transition,
Parr (f romn0 , f romn1 , ton0 , ton1 ) = Nadmitted /Ntotal
where,
   
f romn0 f romn1 f romnoverlap
Nadmitted =
  c0 c1 coverlap
N
Ntotal =
w
For a non admissible transition,
Parr (f romn0 , f romn1 , ton0 , ton1 ) = 0

D. Departure state transition probabilities

The departure probabilities from one state to another state
are proportional to the arrival rates into the state. Therefore,
Pdep (s0 , s1 , d0 , d1 )
Parr (d0 ,d1 ,s0 ,s1 )
= PN P N
Parr (n0 ,n1 ,s0 ,s1 )
n0 =0 n1 =0

6 If a packet cannot be transmitted upon arrival, the system undergoes a

same state transition and the packet is dropped. During a same state transition
the number of packets on the line does not change.