Professional Documents
Culture Documents
There has been rapid growth of wireless telephony in the past few years. In
onjun
tion
with the explosive growth of the Internet, this has led to an in
reasing demand for wireless
data servi
es. In response to this demand, various me
hanisms have been proposed to
support data traÆ
over wireless telephony servi
es. One of the proposed s
hemes to
provide high-speed downstream data a
ess over CDMA is High Date Rate (HDR, see
[2℄).
In this paper, we wish to address the following problems, namely,
(i) How
an multiple real-time data users be supported simultaneously with good qual-
ity of servi
e (QoS) for all users, namely, with pa
ket delays not ex
eeding given
thresholds with high probability.
This work was
arried out when this author was an intern at Bell Labs, Lu
ent Te
h., Murray Hill, NJ
07974
(ii) How
an a mixture of real-time and non-real-time users be supported simultaneously
with real-time users re
eiving their desired QoS and non-real-time users re
eiving
the maximum possible throughput without
ompromising the QoS requirements of
real-time users.
Wireless s
heduling has two pe
uliarities whi
h distinguish it from
onventional wireline
s
heduling. These are
(i) The
hannel here is not perfe
t and is subje
t to errors. This
auses bursts of errors
to o
ur during whi
h pa
kets
annot be su
essfully transmitted on link. The
impli
ation of this is that good s
heduling algorithms need to be
hannel quality
(state) dependent.
(ii) Channel state varies randomly in time on both slow and fast time s
ale.
(a) Fast
hannel variations (due to fast fading) are su
h that states of dierent
hannels
an asyn
hronously swit
h from \good" to \bad" within a few millise
onds
and vi
e-versa. A good s
heduling algorithm should take advantage of this by giving
some preferen
e to a user whose
hannel is
urrently good.
(b) Slow
hannel variation means that the average
hannel state
ondition de-
pends on user lo
ation and interferen
e level. Therefore some users inherently de-
mand more air-interfa
e resour
es than others, even if their data rate requirement
is the same.
The study in this paper
ompares various wireless s
heduling algorithms whi
h expli
itly
use the above information. Our model and the form of QoS for real-time users are essen-
tially same as those in [1℄, where the Modied Largest Weighted Delay First (M-LWDF)
rule was proposed and studied. The most important
ontribution of this paper is the
detailed study of the Exponential rule and the gains a
hieved by the use of this algorithm.
(In [12℄, we also formally show that the exponential rule is throughput-optimal, in the sense
that it makes the queues stable if it is feasible to do so with any other s
heduling rule.)
We also study means by whi
h users
an be guaranteed a minimum throughput, namely
employing virtual token queues for this purpose. Guaranteed minimum throughput is a
QoS notion appropriate for non-real-time users.
1.1. Related Work
In [16℄, optimal s
heduling for a wireless system
onsisting on N queues and a single
server is studied. The arrival pro
ess to ea
h queue are assumed to be and i.i.d. Bernoulli
pro
esses. The
hannel per
eived by ea
h queue is also assumed to be an i.i.d. Bernoulli
pro
ess. The authors show that the poli
y whi
h minimizes the total number of pa
kets
in the system in a sto
hasti
ordering sense is the one whi
h serves the longest queue.
However, if the
hannels are
orrelated (as is the
ase in Rayleigh fading
hannels), then
it is
lear that a
hannel state independent poli
y like this one will perform badly.
In [13℄, the authors study the problem of s
heduling multiple real-time streams with
deadlines, over a shared
hannel. The
hannel is
onsidered to be ON-OFF. It is observed
that a
hannel state dependent version of the earliest deadline date (EDD) poli
y is
not always optimal in the sense of minimizing the number of pa
kets lost due to deadline
expiry. However, for most values of the
hannel parameters that are of pra
ti
al interest, it
is shown that the modied-EDD is optimal. Re
ent work in [18℄ study a dynami
program
approa
h to downlink s
heduling, where for various amounts of
hannel information, the
authors
hara
terize poli
ies whi
h maximize a dis
ounted
ost using a dynami
program.
The authors in [3℄ study the ee
t of the wireless link on the performan
e of trans-
port proto
ols su
h as TCP for various s
heduling proto
ols using a simulation-based
approa
h. They
on
lude that a
hannel-state dependent s
heduler
an lead to signi
ant
improvement in
hannel utilization for typi
al wireless LAN
ongurations. This idea is
further explored in [7℄ in the
ontext of fair queueing.
The rest of this paper is as follows. In Se
tion 2, we des
ribe the system model and
dis
uss the simulation parameters. We then des
ribe various s
heduling algorithms (in
Se
tion 2.1) under
onsideration and the performan
e metri
s (in Se
tion 2.3) that are
used. Finally, in Se
tion 3, we present the simulation results and dis
uss them.
2. DESCRIPTION OF THE MODEL
A very attra
tive feature of HDR is that it enables the use of eÆ
ient s
heduling
algorithms. In this se
tion, we des
ribe the CDMA-HDR system model, and des
ribe
various algorithms used for s
heduling data
ows. Two measures of the merit of these
algorithms are pa
ket delay distributions and throughput. We dis
uss these measures
and des
ribe a means of ensuring quality of servi
e (QoS) guarantees in
onjun
tion with
these measures.
We study a single
ell CDMA system with HDR (see [2℄). HDR is a downlink pa
ket-
data servi
e whi
h o
upies a single
arrier of a CDMA system. Transmissions for dierent
users are time multiplexed, i.e., data is transmitted to one user at a time at the full power
available at the base-station. The
ell serves N mobile users, ea
h re
eiving a data
ow2 .
The base-station
ontains N queues, ea
h
orresponding to a dierent data
ow, and an
asso
iated s
heduler. The s
heduling de
isions are made on
e every \time-slot", whi
h is
as in HDR, 1.667 m se
.
As mentioned before, all users share a
ommon
hannel. However, the quality (i.e., the
fading level) of the
hannel seen by dierent users is dierent, and
hanges asyn
hronously
in time. The
hannel model in our simulation is as follows. Using typi
al HDR and
ell
parameters, and using a
onservative analysis as in [2℄, we
an derive the average fade
level distribution for a typi
al mobile in a
ell. Using this distribution, we pi
k 14 users,
and nd the
orresponding average data rate to ea
h user as in Table 1. Ea
h entry
orresponds to the mean data rate re
eived by the
orresponding user if the base-station
dedi
ated all transmission time to that user. Then, we assume that the
hannels are
Rayleigh fading with the mean being that
hosen above. We next note that HDR does
not support arbitrary transmission rates. In our simulations, we assume that the a
tual
set of available rates is 2 *9.6 kbps, i = 0; 1; 2; : : :3. The a
tual transmission rate is
i
hosen to be the largest available rate su
h that is satised. We denote the a
tual
Eb
maximum rate that the
hannel
an support over a slot as the state of the
hannel at
N0
that slot. Finally, in our simulations, we assume that the speed of the mobiles are 6 mph.
2 This means that the pa
kets of ea
h
ow must be delivered in order.
3 The exa
t rate set in HDR is given in [2℄.
Table 1
Mean throughput supported by the Rayleigh fading
hannels.
User Number \Mean" Rate ( in kbps) i
1 398.72
2 397.14
3 397.14
4 384.91
5 360.88
6 342.00
7 276.01
8 241.43
9 232.18
10 220.60
11 191.24
12 154.27
13 150.72
14 137.80
This
orresponds to a doppler frequen
y of 8 Hz, roughly
hara
terizing how \fast" the
variations of the
hannel quality are.
The s
heduler makes a de
ision to serve a parti
ular queue at the beginning of every
time slot. This de
ision
ould depend on pa
ket delays and
hannel states. We assume
that the s
heduler has (
urrent or delayed)
hannel state information and also the head
of line (HOL) pa
ket delays. On
e a de
ision has been made, the
hosen queue is servi
ed
in the slot at the rate
orresponding to the state of its
hannel.
In the simulations, we assume that N = 14. Ea
h of these users require an average rate
of 28.8 kbps. This
orresponds to the typi
al rate required for streaming audio over the
Internet. The arrival pro
esses to ea
h of the queues are Bernoulli pro
esses with a mean
rate of 28.8kbps, with a pa
ket size being 128 bytes (
orresponding to a HDR pa
ket).
Real-time users like streaming audio will indeed generate a smooth traÆ
, and hen
e, a
Bernoulli model seems reasonable for su
h traÆ
. On the other hand, in our study, a
non-real-time or a very bursty traÆ
is modeled by its \worst
ase", an innite amount
of data to send.
2.1. S
heduling Algorithms
The s
heduling algorithms under
onsideration are the FIFO rule, maximum rate rule
(MAX-RATE), the modied largest weighted delay rst rule (M-LWDF), the proportion-
ally fair rule (PROP-FAIR), and the exponential rule (EXP).
Let us dene
(t) to be the state of the
hannel of user i at time t, i.e., the a
tual rate supported
i
by the
hannel. This rate is
onstant over one slot.
i to be the rate
orresponding to the mean fading level of user i as in Table 1.
W (t) to be the amount of time the HOL pa
ket of user i has spent at the base-
i
station.
The FIFO dis
ipline s
hedules the user whose HOL pa
ket has spent the longest time
at the base-station. Formally, we s
hedule at time t, the user
j = arg max W (t) i
i
This poli
y is oblivious to the
hannel state, and as
an be expe
ted, performs very
poorly. In the results presented in Se
tion 3, we do not in
lude the performan
e of this
poli
y be
ause for typi
al loads supported by other poli
ies, this poli
y renders the system
unstable.
The maximum rate rule s
hedules the user whose
hannel
an support the largest data
rate over the next slot, i.e.,
j = arg max (t) i
i
This
learly requires
hannel state information. We study the
ase where the exa
t
hannel
state is known and also the
ase when only delayed
hannel state information is available.
The modied largest weighted delay rst (M-LWDF) rule has been proposed in [1℄. For
any arbitrary set of
> 0, the rule is given by
i
It has been proven in [1℄ that this rule is throughput optimal. (A poli
y is said to be
throughput-optimal if it satises the property that it renders the queues at the base-
station stable if any other rule
an do so. In other words, it has the largest stable
admission region.) It was also demonstrated by simulation that a \good"
hoi
e of
is i
given by
= , where a > 0; i = 1; : : : ; N , are suitable weights, whi
h
hara
terize the
i
ai
i
desired QoS. This rule tries to balan
e the weighted delays of pa
kets and also tries to
i
Remark. The above rule (1) is not the \true" proportionally fair rule as in [17,8℄. The
true proportionally fair rule is given by
(t)
j = arg max (2)
i
i ~
i
where ~ is the mean rate a
tually given to user i, and measured over a
ertain (relatively
i
long) \sliding window" (see [17,8℄). In the
ase when all users have innite amount of
data to send to (and all users have the same model of
hannel variations but with dierent
average levels, as in our simulations), the rules (1) and (2) are approximately equivalent.
We
hoose the version (1) for our simulations, be
ause in the
ase of data
ows arriving
at nite rates, for some users we have ~ = (i.e. the entire
ow \goes through") no
i i
matter how good their average
hannel
ondition is; as a result, for those users, ~ is not i
a good measure of the a
tual \amount of the air interfa
e resour
e" allo
ated to them.
We note that the PROP-FAIR rule (any version!) does not a
ount for delays of pa
kets
and
an result in poor delay performan
e (see Se
tion 2.3). Further, it
an be easily shown
that this rule is not throughput-optimal.
Study of the exponential rule, brie
y introdu
ed as a heuristi
s in [1℄, is in the
enter
of this work, so it is dis
ussed in more detail in the next Se
tion.
2.2. The Exponential Rule
Let us x any set of positive
onstants
i > 0 and a i > 0, i = 1; 2; : : : N . Then, the
exponential rule is given by
!
ai Wi (t) aW
j = arg max
(t) exp p
i
i i
1+ aW
where aW = 1 P a W (t).
N i i i
2.2.1. Intuition
For \reasonable" values of
and a (this issue is dis
ussed in Se
tion 2.2.3), this poli
y
i i
tries to equalize the weighted delays a W (t) of all the queues when their dieren
es are
i i
large. As we
an see, if one of pthe queues would have a larger (weighted) delay than
the others by more than order aW , then the exponent term be
omes very large and
overrides
hannel
onsiderations (as long as its
hannel
an support a non-zero rate),
hen
e leading to that queue getting priority. (We note that the aW term in the exponent
an be dropped without
hanging the rule as it is
ommon for all queues. This is present
only to emphasize the motivation for pthis rule.) On the other hand, for small weighted
delay dieren
es (i.e., less than order aW ), the exponential term is
lose to 1 and the
poli
y be
omes the proportionally fair rule. Hen
e, this poli
y gra
efully adapts from a
proportionally fair one to one whi
h balan
es delays. The fa
tor 1 in the denominator of
the rule is present simply to prevent the exponent from blowing up when the weighted
delays are small. We dis
uss the merits of this rule in greater detail in Se
tion 3.
2.2.2. Throughput-Optimality
In our
ompanion paper [12℄, we prove that the exponential rule is throughput optimal.
(We remind, this means that it renders the queues at the base station stable if any other
rule
an do so.) The assumptions on the
hannel and arrival pro
esses are quite general.
We assume that the aggregate
hannel pro
ess (in
luding states of all individual
hannels),
is Markovian with a nite number of states. Note that the individual
hannels need not
be independent. The aggregate input pro
ess is assumed to be an ergodi
(dis
rete time)
Markov
hain.
The proof of the throughput-optimality uses
uid limit te
hnique [10,5,4,14,6℄, along
with a separation of time-s
ales argument. We refer reader to [12℄ for the details.
2.2.3. Choi
e of Parameters
As in M-LWDF, the
hoi
e of
is given by
= . The values of a ; i = 1 : : : N ,
i i
ai
i
determine the tradeo between redu
ing weighted delays and being proportionally fair
i
when delays are small. The QoS requirement would be of the form
Pr(Wi > Ti) Æ i
where W is delay en
ountered by a typi
al pa
ket of user i, T is the maximum delay that
i i
user i
an tolerate (measured in slots) and Æ is the largest probability with whi
h the
i
system is allowed to violate the delay requirement. For the system we are interested in, a
typi
al requirement would be a violation probability of 10 2
orresponding to a maximum
delay of the order of a se
ond. As play-out buers for streaming audio over the Internet
are suÆ
iently large, the above values are a reasonable requirement. log( )
A rule of thumb for
hoosing a whi
h works in pra
tise is a =
i i . This rule is
Æi
suggested by large deviation optimality results of [15℄. We note that the performan
e
Ti
of the rule is robust with respe
t to moderate
hanges of T and Æ . Simulation results
i i
exploring this aspe
t are available in [11℄.
2.3. Performan
e Metri
s
In this se
tion, we dis
uss the performan
e metri
s we employ to evaluate the s
heduling
algorithms. These are the delay
riterion and the average throughput. One or the other
is more important depending on the type of the traÆ
.
For real-time traÆ
, a good measure of performan
e is the delays pa
kets in
ur at the
base-station. We assume for simpli
ity that all users need the same delay QoS, i.e., they
all have the same T and Æ . A good s
heduling algorithm should keep all delays below
i i
T with high probability, whi
h is a
hieved roughly when the delays are kept
lose for
i
all users. In the setup we study, the worst user (the mobile with the worst
hannel)
has the average
hannel quality mu
h worse than the best one. From Table 1, the best
user
an support a mean rate of 398.72 kbps (if it would be the only user of the
ell)
while the worst one supports only 137.8 kbps. However, we want to support 14 users
simultaneously, ea
h with a mean rate 28.8 kbps, leading to a total rate of 403.2 kbps.
A naive s
heduling algorithm whi
h does not depend on the
hannel state will nd it
impossible to a
hieve this obje
tive. However, a smart algorithm whi
h favors users with
relatively good
hannels would hopefully be able to a
hieve this obje
tive.
If a system with real-time users only is stable, then
learly, the long term throughput
is simply the arrival rate. As we remarked earlier, the M-LWDF and the EXP rule are
optimal in this regard, i.e., they
an support the largest set of available rates. The proofs
of this
an be seen in [1,12℄. Finally, we
omment that s
heduling algorithms whi
h keep
the delays of all the users about the same and keep them all reasonably small are superior
to those whi
h may have better delay tails for one of the users but have very bad delays
for other users.
Another important measure is the long term throughput available to ea
h of the queues.
For non-real time data, this is the main measure of interest. We would like to give
minimum throughput guarantees to these users.
Suppose we have a mixture of
oexisting real-time and non-real-time users in the
ell.
In this
ase, we want to serve all the real-time users with low
ow delays and provide
minimum rate guarantees to non-real-time users. Moreover, we want leftover
apa
ity to
be allo
ated to non-real-time users in a proportionally fair manner. We show that this
an be a
omplished using token based versions of our s
heduling s
hemes.
2.4. Token Queues for Minimum Throughput Guarantees
The M-LWDF and EXP rules, as they des
ribed above, make s
heduling de
isions based
on the a
tual pa
ket delays (in addition to the
hannel
onditions). If providing a
ertain
minimum throughput to a
ow is a goal, those rules
an be modied as follows. Suppose,
asso
iated with ea
h user, there is a virtual token queue into whi
h tokens arrive at a
onstant rate r , the desired throughput of user i. (I.e., a
ounter representing the token
i
queue is in
remented by r t at regular time points, t time units apart.) Let us dene
i
V (t) to be the delay of the head of line token in the user i token queue. Note that we
i
do not need to a
tually maintain the token delays. As the arrival rates of tokens are
onstant, V (t) = ( ) , where Q (t) is token queue length (a
ounter value) at time t.
i
Qi t
i
Then, we use the M-LWDF and EXP rules with W (t) being repla
ed by V (t). After
ri
i i
the servi
e of a (real) queue, the number of tokens in the
orresponding token queue is
redu
ed by the a
tual amount of data transmitted. (\Pathologi
al"
ases are treated as
follows. If a user is s
heduled, and does not have enough pa
kets to send, the token queue
is depleted by the amount as if it had enough data to send. If user should be s
heduled,
but has no data to send, its token queue is depleted by the
orresponding amount, and
s
heduler
hooses the \se
ond best" queue to serve, and so on. If a token queue is shorter
than the number of tokens that needs to be removed, it is set to 0.) To in
rease the
algorithm's robustness, for real-time users, we will set the token rates to be slightly larger
(by 2%) than the desired throughput.
A token queue is similar to a leaky bu
ket regulator followed by a real queue based
s
heduler. However, there are some dieren
es. With a good s
heduling algorithm, a
token queue me
hanism
an allow a user to utilize more than the required minimum rate
if all the token queues are small. Su
h an ee
t is possible with leaky bu
ket regulators
if the regulator marks pa
kets as high and low priority and the s
heduling algorithm
operates on the high priority queue lengths and pa
ket delays. However, this
ould lead
to out sequen
e delivery for the low priority pa
kets. This would not be so with the token
queue me
hanism.
3. SIMULATION RESULTS AND DISCUSSION
In this se
tion, we present simulation results for the system des
ribed in the previous
se
tion. As des
ribed earlier, we
onsider a CDMA/HDR
ell with 14 users. The
han-
nels are Rayleigh fading, whi
h have been implemented using Jakes's implementation of
Clarke's fading model.
3.1. Real Queue based S
heduling
In Figure 1, we plot the delay distribution tails for the best and worst users(i.e., users 1
and 14 in Table 1). The arrival pro
ess to ea
h queue is a Bernoulli pro
ess with a mean
rate of 28.8 kbps. We assume that the s
heduler has knowledge of the
urrent state of the
hannel. We observe that for the best user, the MAX-RATE rule provides the best delay
distribution tails. However, for the worst user, the queue be
omes unstable, leading to
Pr(Wi > d) = 1 for all d > 0. This
an be understood from the fa
t that at ea
h slot, this
algorithm s
hedules the queue whose
hannel
an support the maximum rate. Therefore,
most of the time, this leads to the user
orresponding to the better
hannel utilizing a
slot whenever it has pa
kets. However, this greedy s
heme leads to suboptimal utilization
14 Users without Token Queues
0
10
EXP
M−LWDF
MAXRATE
PROP FAIR
−1
10
−2
10
Pr(Delay > x)
−3
10
−4
10
−5
10
0 500 1000 1500 2000 2500 3000
delay (m sec)
Figure 1. Real queue based s
heduling. Data rate: 28.8 kbps per user. Delay tails of the
best and worst users are plotted. The s
heduler has exa
t
hannel knowledge.
50
Acheived rate − Desired rate (kbps)
40
30
20
10
−10
0 2 4 6 8 10 12 14
User
Figure 2. Token Queue based s
heduling. Data rate: 28.8 kbps per user. User 1 has
innite ba
klog.
−1
10
−2
10
Pr(Delay > x)
−3
10
−4
10
−5
10
0 500 1000 1500 2000 2500 3000
delay (m sec)
Figure 3. Token Queue based s
heduling. Data rate: 28.8 kbps per user. User 1 has
innite ba
klog. The delay distribution tails of users 2 and 14 are plotted.
ompared to the other rules. This observation
an be explained by the fa
t that when
token queues are small, the EXP rule is basi
ally the same as the PROP FAIR rule.
However, when delays start be
oming unbalan
ed, the exponential term balan
es the
delays. Further simulation results for other systems (see [11℄) also bear out our previous
observations. These results indi
ate that we
an indeed support 14 users with the required
rates and hen
e, as dis
ussed in Se
tion 2.3, supporting more than any single users' average
hannel throughput!
4. CONCLUSIONS