gl-1

a,-1

under high load when all input queues become occupied. The algorithm performs efficiently and fairly under low and high load and is simple to implement. One drawback of the algorithm is that it requires more than one iteration in any slot time. A modification to the IRRM algorithm called SLIP-IRRM provides a use factor asymptotic to 100% in a single iteration.

.
Acknowledgments: We would like to thank R. Edell and T. Anderson for their helpful advice and comments. I 2 October 1993 0 IEE 1993 Electronics Letters Online No: I9931420 N. McKeown, P. Varaiya and J. Walrand (Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720, USA)

input output Fig. 1 One iteration of the three phase iterative round-robinscheduling algorithm are occupied, it can take min(M,N) iterations to successtully schedule the switch if the schedulers are initially aligned synchronised. However, in successive cycles, a perfect match will occur in a single iteration provided that all input FIFO queues are still occupied. This is the reason for the asymptotic 100% use factor under high load independent of the traffic statistics. We select a fixed number of iterations, T, where 1 < T << min(M,N). Under high load the schedulers will become misaligned over several time slots. Once they have become misaligned, matches will continue to be perfect so long as the input queues are occupied. The basic IRRM algorithm will not perform well for a single iteration, because the basic algorithm does not allow the schedulers to become misaligned. However, if we make the following small change to the basic algorithm, the maximum achievable use factor becomes asymptotic to 100%. The SLIP-IRRM algorithm is identical to IRRM with the following condition added to phase 2 above: (2) ... Pointer g, is incremented if and only if the grant was accepted.

References
1
KAROL, M., HLUCHYJ, M., and MORGAN. s.: Input versus output queueing on a space division switch, IEEE Trans., 35, (IZ), pp. 1347-1 356 OBARA, H.: An efficient contention resolution algorithm for input queueing ATM switches, Int. J. Digital & Analog Cabled Systems, 1989, 2, (4), pp. 261-267 O B A R A , H . : Optimum architecture for input queueing ATM switches, Electron. Lett., 1991, 27, pp. 551-557 OBARA, H., OKAMOTO, S., and HAMAZUMI. Y.: Input and OUtpUt queueing ATM switch architecture with spatial and temporal reservation control, Electron. Lett., 1992, 28, pp. 22-24 KAROL, M., ENG, K., and OBARA, H.: Improving the performance of input-queued ATM packet switches. INFOCOM 92, 1992, pp.

2
3

4
5

T
25

110-115 6 ANDERSON, T., OWICKI, s., SAXE, J., and THACKER, c.: High speed

I R R M . ~ ~ ~ SLIP-IRRM iteration j

x 15
3

i9
k 5

switch scheduling for local area networks. Roc. Fifth Int. Conf. on Architectural Support for Programming Languages and Operating Systems, October 1992, pp. 98-1 IO

Statistical modelling of spectrum occupancy

A.J. Gibson and L. Arnett
Indexing terms: Radiocommunicatiom,Modelling, Frequency allocation, Markov processes A novel method of modelling spectrum occupancy is identified. Good agreement is demonstrated between a conceptual formula for new two-dimensional transistion probabilities and estimates from experimental data. Application to radio system design and frequency management is explained.

O i

40 60 80 100 use, % Fig. 2 Average bufler occupancy in cells against percentage use factor for iterative round-robin matching algorithm

I I 0

20

Switch has 16 inputs and 16 outputs; traftic model is independent Bernoulli arrivals with destinations uniformly distributed over all outputs
(ii) Performance of iterative round robin matching; Fig. 2 shows the use factor against delay for both IRRM and SLIP-IRRM for a 16-input 16-output switch. IRRM was simulated for one, two and four iterations. For SLIP-IRRM we can see that the use factor is once again asymptotic to 100% because of the forced misalignment of the schedulers introduced by the modification. In addition to the high use factor of this algorithm, IRRM and SLIP-IRRM will always allocate bandwidth fairly between contending inputs. This is because of the deterministic nature of round-robin scheduling. Conclusion: We have introduced a new algorithm called IRRM for scheduling cells in an input-queued cell switch. The algorithm exhibits the interesting property of asymptotically 100% use factor

Introduction: Spectrum occupancy data are pertinent to radio system design and frequency management strategies. However, detailed measurements and analyses intended to quantify the performance of one particular system usually cannot be applied directly to other cases. An alternative is that data analyses should be directed to developing a statistical model of more general applicability. Existing mathematical models [I, 21 are essentially for one variable (e.g. congestion), whereas this Letter proposes that three variables are required. The first stage is to model the occupancy of a given frequency band at a given time and place in terms of parameters from which can be calculated quantities required for system design such as clear channel duration and its bandwidth dependence. The second stage, not discussed in detail here, will be to model those parameters as functions of time, frequency, geographical location, etc. Conceptual model identification: Traditionally, the occupancy of a single channel has been defined as the unconditional probability q that the measured signal strength exceeds some predetennined threshold. If X, denotes whether the ith measurement is below threshold
q = P[,, = 0 1

p = 1 - q = P[z; = 1 1

(1)

ELECTRONICS LETTERS 9th December 1993

Vol. 29

No.25

2175

Averaged over a band of channels, q and p have been called the 'congestion' [I] and 'clearness' [2], respectively. Spaulding and Hagn [3] suggested that n successive values of x, for a single channel could be modelled as a first-order Markov chain. The usefulness of this concept is that, given p and n, one further parameter, the conditional probability

x = P[s,= 112,-1

=1 1

(2)

is sufficient to determine all the probability properties of the channel. The other three transition probabilities can be expressed in terms of p and A, and the joint probability distribution [4] is PIX1 = q , X 2 = I2,.. .,xn = I n ]= xr(1 -x)z(s-r)-t where
(1-2p+Xp)"-'-2"+t+'

S-r

-(n-Z-s+t)

(3)

I,
11 12 13
1 L

I
15

,=c2.-lz. s=c.r*
n 2=2

frequency, MHz

t=x:l+sn

(4)

rn

*=I

Fig. 1 Example of romparison of values of A estimated directly from dota and calmlared using eqn. 7
-A'=u!u ----- A =

Laycock et al. [SI indicated that the occupancy of n, successive channels in a particular allocation in a single scan of the spectrum could also be modelled as a two-state Markov chain; the transition probability corresponding to eqn. 2 can be denoted by p, and the joint probability distribution analogous to eqn. 3 is useful in determining the probability properties of a single scan. Combining these suggestions, denoting the state of theflh channel on the ith Scan by A',, we identify eight two-dimensional transition probabilities denoted by
dblm = P [ z , = klz(,-l,3

(A + p + h p - p ) / ( l - p )

spectrum, an average over about 60 days of the values o f h ' estimated directly from the data, together with the values obtained indirectly using eqn. 7. The agreement in Fig. 1 and similar plots for other parts of the spectrum is encouraging, but it is intended to carry out additional tests [7] to validate the model.
Application: A number of formulas relevant to system design can be derived using the joint probability distribution. For example, applying eqn. 3 to a single scan, the probability of no channel below threshold is

= 1,1,13-1) = m] k l , m = 0 , l

(5) The conditional probability that a channel is helow threshold (clear), given that it was clear on the previous scan, and given that the adjacent channel on the same scan was clear, denoted by

A = dl11= P[z,, =

=l

. ~ ~ ( ~= - l 1 1 l

(6)

P ( 0 )=

(1 - 2p
Q " '

+ ppp)"?-'
-2

is a parameter of interest in calculating durations of composite channels. We have derived [6] expressions for the other seven transition probabilities defined by eqn. 5 in terms of p , A, p and A, and shown that, in the first-order approximation

A=

x+p

xfi - p

1-P

17)

The probability that m adjacent channels are below threshold is pp"-' and the probability that they remain below threshold for one sampling interval is ?A"-'. Thus, a second-stage model of three independent variables (p, h and p normalised to the sampling intervals), rather than a single variable (p or q), enables many more quantities of interest to be derived.
0 Crown copyright1993 Electronics Letters Online No: 19931436

This can also be written in the form

15 October 1993

A = A + (1 - X ) p

(8)

where p is the intra-channel correlation coefficient defined in [SI. 'Thus, the model is specified by three independent parameters p. h and IL.
Estimafion of model parameters from data: In the one-dimensional case, the maximum likelihood estimators of p and h have been shown [4] to be unbiased and asymptotically efficient. However, in the two-dimensional case, we use the relative frequency estimates

A.J. Gibson (Radiocommunications Research Unit, R25. Rutherford Appleton Laboratory, Chilton, Didcot, Oxon. OX11 OQX, United Kingdom)
L. Arnett (DRA Malvern, St. Andrew's Road, Great Malvern, Worcs WR14 3PS. United Kingdom)

References
POOLE. c.R., LAYCOCK. P.J., CHAN. s.K., and RAY,A.R.: 'A model for HF spectral occupancy'. Proc. 5th Int. Conf. on HF Radio Systems and Techniques, 1991, (Edinburgh), IEE Conf. Pub. 339, pp. 332-336 MUTERLARD. c., and CARATORI. J.: 'Time modelling of HF interferences'. Proc. 5th Int. Conf. on HF Radio Systems and Techniques, 1991, (Edinburgh),IEE Conf. Pub. 339,pp. 343-348 SPAULDING, A.D., and HAGN. G.H.: 'On the definition and estimation of spectrum occupancy', IEEE Trans., 1977, EMC-19, pp. 269-280 KLOTZ. J.: 'Statistical inference in Bernoulli trials with dependence', Annals on Statistics, 1973, 1, (2). pp. 373-379 LAYCOCK. P.J., MORRELL, M., OOTT. G.F., and RAY, A.R.: 'A model for HF spectral occupancy'. Proc. 4th Int. Conf. on HF Radio Systems and Techniques, 1988, (London), IEE Conf. Pub. 284, pp. 165-1 71 GIBSON, A.J.: 'A statistical model of spectrum occupancy'. NRPP Research Note No. 142, RCRU, RAL, September 1993 ANDERSON, T.W., and M O D M A N . L.A.: 'Statistical inference about Markov chains', Ann. ofMath. Stat., 1957, 28. pp. 89-110
GOTT.G.F.,

(9)

where

and

Validation: Estimates of p , h, p and A were obtained for blocks of 25 2kHz channels and 15 scans, using eqns. 9-11 on real data. Fig. 1 shows, for 100 such SokHz blocks spread over 5MHz of the

2176

ELECTRONICS LE77ERS

9th December 1993

Vol. 29

No. 25

-_