You are on page 1of 5

Asymptotic Analysis of Proportional Fair Algorithm

Jack M. Holtzman
Qualcomm, Inc.
5775 Morehouse Drive, San Diego, Ca 92121-1714
jholtzma@qualcomm.com

Abstract fairness - a user which does not get transmitted


for too long gets its priority raised.
It was previously shown (under simplifying
assumptions) that this algorithm gives equal In [5] we presented a result which yields a simple
power and time to users who only differ in the interpretation of this algorithm showing how it
distance from the BS, their fading characteristics provides a throughput leveling and an appealing
being the same. We give here a more general tradeoff between throughput and fairness. We
result for two classes of users with different provide here an asymptotic analysis of the
fading characteristics. All other things being algorithm for two classes of users, which includes
equal, the user class with more fading variability the interpretation of [5] as a special case. All
gets more throughput with a lower (but not much other things being equal, the user class with more
lower) fraction of time transmitting. We also fading variability gets more throughput with a
demonstrate a separability property - the effects lower (but not much lower) fraction of time
of variablity and distance from the base station transmitting. We also demonstrate a separability
are separable. property - the effects of variablity and distance
from the base station are separable.
I. Introduction
There are a number of other papers on downlink
In [ l ] - [3], it is shown how knowledge of the scheduling, e.g., [6] which has further references.
channel can increase CDMA capacity by
transmitting mostly when the channel is good. An 2. Proportional Fair Algorithm
issue that arises in cellular CDMA is that of A simplified version is as follows: Let
fairness because users nearer the base station P(t) = available power at time t for use by K
(BS) would be favored in such approaches. There delay-tolerant data users.
is thus a tradeoff between total throughput and Tk(t) = throughput of user k over a time window
fairness among users. up to timet, k=1, ...K
The user with highest priority in a time slot
An algorithm based on priority given just by the defined by
CA would always give all the power to the user
(c/z)k (t>
close to the BS with the best channel, maximizing (1)
system throughput but also being unfair to users Tk ( t )

far from the BS. An algorithm was recently transmits with power P(t).
introduced which strikes a good compromise The throughput monitoring displayed in (1)
between throughput and fairness [4]. This introduces fairness -- a user which does not
algorithm (called a proportional fair algorithm) transmit for too long because of a relatively low
includes throughput monitoring which introduces C/I gets its priority raised by the low
denominator.
0.7803-7244-1/01/$10.00 a2001 IEEE

F-33
Remark on CA. It is implicitly assumed that the All the {bl,j( t ) ) and {b2,j( t ) ) are iid within each
supportable data rate is proportional to CL, class and independent across classes (also
which is usually a reasonable CDMA independent across time slots). This is in contrast
approximation for not too high CA. It is not to [5] where & users’
l random components were
accurate when multiple modulations or codings iid’s.
are used. This will be generalized in Section 4.
Now consider the window of observation for
3. Simple Interpretation of Algorithm from throughput monitoring +- so that TI and T2
PI are considered converged to their stationary
Let values. We are interested in determining the ratio
of throughputs,
( C / Z ) , ( t ) = a ,. b , ( t ) , k=1,...,K
r = -T
a,= distance dependent component of C/I for
user k T2
A user from Class 1 is chosen for transmission if
b , ( t ) = random component of C/I for user k,
including Rayleigh or Ricean fading. XI
->- x,
Assumution I: the b, (t) are iids among the users Tl T2
where
and from time slot to time slot, and they are all
independent of P(t)
Remark. It is assumed that there is one slot for
each user’s transmission. This may not always be
the case - lower data rate users might use more
slots.
Assumution 2: Rate is (approximately) linear Then, since the transmitted rate of a class 1 user
with power (continuous with no limit). is proportional to x,,the total mean throughput
Tk = stationary throughput of user k (with the of class 1 users is proportional to
throughput window W + a; in actuality the
window would be chosen to do some averaging
and short enough to control delays).

It is shown in [5] that, under Assumptions 1 and and a similar expression for class 2. This leads to
2, Tk/ a , is the same for all k, and all users get the ratio of total class throughputs:
transmitted with the same asymptotic fraction of
time and with the same average power.

4. Analysis for Two Classes of Users which is a fixed point equation

Consider two classes of users defined by r =f(r)


The numerator and denominator in the fixed point
equation are evaluated as:

F-34’
and
00 - al = 1, a2 = 10,N1= N2 = 5, K :0+-10
(linear,
NI JPI (4)&I j.2 P2 (x2 )dr, not dB)
0 xzlr
Note that in Figures 3 and 4, the result of [5] is
where the pi@) are the pdf's. (The properties of recovered at K=O (see the Remark preceding the
the order statistics are given in 173). numerical example). That is, the throughputs are,
in that case, in proportion to the average a ' s .
Remark on uniqueness of fued point, Since
f ( r )is positive and montonically decreasing for 4. Generalization
positive r , there is a unique fixed point for
positive r. This is consistent with the uniqueness One direction of generalization is to no longer
of the optimum to the concave optimization aSsume that the supportable data rate is
problem in [4]. proportional to the UI,but rather a function of it

Remark on the special case bl(t) and b2(t)


being identically distributed. It is easily shown
that the solution to the fmed point equation is Then a user's priority is

which is, in fact, the actual proportional fairness


for N, = N 2 as found in [5]. algorithm. Then, the fixed point equation
becomes
Numerical examdes

The first class is Rayleigh distributed and the


second class is Ricean. For the Ricean pdf, we
use p ( x ) as given in (2.44) of [8]:
It was shown above (under simplifying
assumptions) that this algorithm
Gives equal power and time to users who only
differ in the distance from the BS, their fading
characteristicsbeing the same.
0 Gives larger throughput to users whose
variability is greater than others' variability,
where here K is the Ricean factor (linear, not in with their distances to the BS being the same.
a), the ratio of powers of specular and scattered The users with greater variability use a lower
components and a, is the average power. (but not much lower) fraction of time
transmitting.
Results are shown in Figures 1 and 2 for
Here, we show that these effects are separable.
a1 = a2 = 1, N1= N2 = 5, K :0-+10(linear, That is, the combined effect is the superposition
not dB) of one effect upon the other, with each
d e t e d n e d independently. Recapitulating the
and in Figures 3 and 4 for results,

F-35
Note that the second numerical example
When b,(t) and b2(t) are identically (al =I, a2 = i o ) displays this separability.
distributed, 5. Concluding Remark
T --
1 -a1
The results use a number of simplifying
=2 a2
assumptions (including no prediction errors and
and the two user classes get the same fraction of rate quantization, and always full buffers) and
time transmitting. should be regarded as first approximations to
When a, = a 2 , the user class with more provide insight and to help interpret more
variability gets more throughput (and a little accurate simulations. Comparison with some
smaller fraction of time transmitting). simulations show that, although there are
numerical differences depending on specific
To show separability of these effects, the fixed parameters, the analysis well characterizes the
point equation is results qualitatively.

ACKNOWLEDGEMENT

The comments of Boris Tsybakov are


appreciated.

REFERENCES

[ l ] A. Goldsmith and P. Varaiya, “Increasing


spectral efficiency through power control,”
ICC’93.
[2] R. Knopp and P. Humblet, “Multiple-
or with accessing over fi-equency-selective fading
t ’ = r -a2, channels,” PIMRC’95.
[3] S.W. Kim and A. Goldsmith, “Truncated
a1
power control in code division multiple access
the fixed point equation becomes communications,” Globecom’97.
[4] D. Tse, “A proportional fair forward-link
scheduling algorithm,” to appear.
[5] J.M. Holtzman, “CDMA forward link
waterfilling power control,” Proc. VTC2000-

$,,*,,la,(OI
max .-
I j=L**&{h,w)>(f)-l
max
{e(+!
]=~..N,
Spring, pp. 1663-1667, May, 2000.
[6] N. Joshi, S.R. Kadaba, S . Patel, G.S.
Sundaram, “Downlink scheduling in CDMA data
which is independent of (a, , u 2 ) . Hence, it is networks,” ACM Mobicom2000.
seen that first the result for variability can be 171H.A. David, “Order Statistics,” 2d Ed., Wiley,
obtained with arbitrary (ul ,a 2 ) , and then the 1981.
[8] G.L. Stuber, “Principles of Mobile
distance effect can be obtained from Communication,” Kluwer, 1996.
r=r’-a1
a2

F-36
0.16
J.52 le6 ,O. 152 , I
1.55
hi 1.5
1.45
m'
a 1.4
c 1.35
Ratiq 1.3
8
).
1.25
1.2
.-0
0-
1.15
d 1.1
1 .05 Lo*L0*08b 2 4 6 8 !O
J , 1 ,a Ki J0
Ricean factor, K (linear, not dB)
,a Ki J0 w Tl/T2
Ricean factor, K (linear, not dB)
Tl/T2 Figure 3-Throughput Ratio vs. Ricean Factor
K, al=l, a2=10. Nl=N2=5
Figure 1-Throughput Ratio vs. Ricean Factor
K, al=a2=1, Nl=N2=5

-8
1- ,
,a K! JO
LE 2 4 6 8 10
Ricean factor, K (linear, not dB)
,a Ki J0 w fractionoftime
Ricean factor, K (linear, not dB)
w fractionoftime Figure 4 - Fraction of Time that User of Class 1
Transmits, a k l , a2=10, Nl=N2=5
Figure 2 - Fraction of Time that User of Class 1
Transmits, al=a2, Nl=N2=5

F-37

You might also like