Вы находитесь на странице: 1из 5

Weighted Round Robin Scheduling Strategies in

(E)GPRS Radio Interface

A. Kuume, A. P. Miettinen
Nokia Networks
Espoo, Finland
antti.kuume, antti.p.miettinen, both@nokia.com

Abstract-Different types of data services have various The objective of this study is to examine the benefits of
requirements for delay and throughput. When a service mix QoS differentiation in (E)GPRS radio interface. Three different
containing both delay sensitive and insensitive types of services is services, streaming media, web browsing and messaging are
offered over a shared channel, it is very beneficial to prioritize presented. With the term 'traffic mix' we refer to proportional
the resource usage of each service. In this paper an analytical volumes of each service with respect to each other. Each
model is presented for studying weighted round robin queuing in
service has its own requirement for throughput to provide
(EDGE) General Packet Radio Service radio interface. A quickly
satisfactory QoE. In a benchmark case the network does not
computable model allows a large number of service mixes to be
apply QoS differentiation. WRR scheduling is introduced as a
studied. The model yields gain achieved by prioritization in terms
method to perform the prioritization. Three different strategies
of maximum carried load with given end user satisfaction
for setting WRR queue weights are proposed. First strategy
criteria. It is shown that according to the model, using weighted
('static') is to keep the queue weights fixed irrespective of the
round robin scheduling brings a notable increase in the
maximum carried load. However, tuning the queue weights either
changing traffic mix. In second strategy ('dynamic') the queue
based on statistical data on average mix of different services or
weights are tuned based on a long-term average traffic mix in
dynamically taking into account the current instantaneous active the network. The third strategy ('adaptive') is to tune the queue
connections do not increase maximum carried load significantly. weights according to the service types of current active
connections in the cell. The objective is to compare these three
different queue weighting strategies among each other and also
1. INTRODUCTION against the benchmark case. In the benchmark case there is no
The term quality of service or QoS is widely used in prioritization in the network at all.
literature in connection with packet switched networks. In this The rest of the paper is structured as follows. In section II
paper with QoS differentiation we mean prioritization of some of the earlier work on the topic is studied and an example
certain packet flows over others in the network. End user QoS differentiation mechanism in (E)GPRS radio interface is
observed QoS is referred to with the term quality of experience, briefly explained. Section III introduces the model and how it
QoE. Objective of QoS differentiation is to provide adequate, has been used in this study. Results are explained in section IV
but not excessive, QoE to all users by sharing the resources and finally conclusions are drawn in the last section.
unequally.
QoS differentiation in General Packet Radio Service II. QoS DIFFERENTIATION IN (E)GPRS RADIO INTERFACE
(GPRS) network (or its enhanced version, EGPRS) can be
enforced for example based on service used, current radio link A. Earlier work
conditions, user subscription class, or a combination of these. In [1] simulations have been run to study weighted fair
On service based QoS differentiation more network resources queuing in a cellular network with bursty data. Users have been
are given to users whose application require high throughput or divided into two distinct priority classes. It has been shown
low delay to reach adequate QoE. This paper considers only how weighted fair queuing can effectively provide significant
service based QoS differentiation. differences to high and low priority users only when the
Weighted round robin (WRR) scheduling refers to network is severely loaded. The focus in [1] is to quantify
treatment of packets (or frames, blocks, etc.) where connection differences in throughput that different priority classes can
i gets transmission tum for ni packets during a period where achieve. The QoS differentiation has been based on user
total number of subscription type, and everyone is using the same service or
application.
(1)
Uplink scheduling in GPRS has been studied in [2]. One of
the introduced methods has been round robin queuing. The
packets are transmitted, where i is running over all connections focus is not on QoS differentiation but more on overall
and ni is the WRR queue weight for connection i. Number of performance. In [3] the priority has been based on block size.
user data bits per packet is not considered. This approach is seen to lead to improved spectral efficiency.
In both [1] and [3] the proportion of traffic volumes in different

0-7803-8521-7/04/$20.00 © 2004 IEEE 3155


priority classes (=traffic mix) has been fixed. Radio network throughput is 10 kbps, the throughputs would be on average 8
has been modeled in detail with 75 cells in [4]. Maximum load kbps and 2 kbps for TBFs A and B respectively.
with given user satisfaction criteria and traffic mix has been
In an arbitrary case of n TBFs sharing one time slot, and
obtained. However due to complex modeling of the radio
assuming constant total throughput for the slot, a following
network and mobility of the users has made it possible to study
formula yields the throughput offered for TBF k on that slot:
only few different traffic mixes. Also in [5] QoS differentiation
with weighted round robin is studied, but the main focus in the 1
paper has been in balancing the resources between circuit
switched voice calls and packet switched data calls.
Rk =
SSSk R (2)
tot'
1
n
The aim in this study is to develop a model that is so simple
and quickly computable that it allows studying of several L-SSS;
;=1
traffic mixes. The aim is to find out if changing traffic mix
indicates that WRR queue weights should not be fixed but where SSSk denotes the value of SSS parameter for TBF k
evolving with the traffic mix. A simple model is believed to be (similarly for i) and Rtot is the total throughput provided by the
sufficient since only relative comparisons are made. We time slot. Note that Rk is the throughput offered by the
estimate that effect of user mobility and protocol modeling
scheduling algorithm for TBF k. The TBF might not use all this
have only minor effects on gain achieved with different queue
capacity. Thus the real throughput of the TBF k is obtained
weights in WRR method.
from

B. (E) GPRS radio interface Rk = min {Rk ,R::ax }, (3)


(E)GPRS is bound to 8 time slot structure of GSM. Each
time slot can send a radio block once every 20 ms. Number of where R::ax is the maximum throughput usable by TBF k. The
user bits in each block varies according to coding scheme used.
There are four coding schemes (CS-l ... CS-4) specified for difference between Rk and Rk is then available to be shared by
GPRS and nine modulation and coding schemes (MCS- other TBFs, according to (2).
1 ... MCS-9) for EGPRS. The (modulation and) coding scheme
used depends generally on radio conditions. The receiver
acknowledges correctly received blocks (there is also an III. THE MODEL
unacknowledged mode of operation). Description about
(E)GPRS radio interface is presented for example in [6]. A. The Model Assumptions
Assume that radio conditions are equal for all terminals.
A physical connection between the sender and the receiver
Then given the number of TBFs sharing a time slot and their
is called a temporary block flow. Each TBF sharing the same
SSS values, (2) can be used to calculate the individual
time slot take turns in transmitting their radio blocks. In case
throughput for each TBF on that particular time slot. Generally
there is no QoS differentiation in the network, two TBFs
however each cell has several time slots available for GPRS
sharing a time slot would receive one block on average once
use. A terminal might be able to use for example three time
per two block periods, i.e. once per every 40 ms. Thus the total
slots simultaneously.
throughput provided by the time slot is shared equally between
the two TBFs, assuming that on average same number of Assume three terminals, each using three time slots in
retransmission are needed and same (M)CS is used. downlink and a cell with five slots available for GPRS. Total
throughput of five time slots (Rtot) is known if we know the
c. QoS differentiation implementation example (equal and stable) radio conditions of the terminals. However
the three phones using three slots each can't be divided to the
One possibility to implement QoS differentiation is to
total five slots of the cell so that each TBF would be in equal
prioritize some TBFs over others by giving them the
position. In this example one TBF gets necessarily allocated to
transmission tum more often. In this study it is assumed that
time slots, which are all shared between some other TBF. The
each TBF has associated with it a parameter called scheduling
two other TBFs get one slot of their own. Fig. 1 illustrates the
step size (SSS), ranging from 1 to 12. After getting to transmit
fairest allocation. TBFs 1 and 3 both get one slot of their own
a radio block, that TBF adds to its existing priority counter a
while TBF 2 shares all its slots with at least one other TBF. In
value equal to TBF's SSS. The next block period is to be used
an example like this, (2) would have to be applied individually
by the TBF with the smallest priority value. This is a WRR
for each time slot. Then the contribution of each slot to the
scheduling system with each TBF having a weight of l/SSS,
throughput of every TBF would be summed up.
where SSS is specific for each TBF.
Assume TBFs A and B are competing of the resources of
the same time slot, and their SSS values are 3 and 12
respectively. Then TBF A gets on average a throughput, which
is 4 times (12/3) higher than the throughput of TBF B. The
numerical throughput value depends on the total throughput Figure l . Example allocation of TBFs 1, 2 and 3 to five time slots.
provided by the time slot. This depends on coding scheme used
and the number of retransmissions needed. Assuming the total

0-7803-8521-7/04/$20.00 © 2004 IEEE 3156


It requires some computation to actually allocate the time (5)
slots for each TBF and apply (2) separately for each slot. In a
real network implementation the time slots chosen generally
depends on several issues. Also reallocations can be made C. The Model
periodically or in an event based manner. In the model this The network consists of Ne identical cells, each having six
issue is omitted and the error in the throughputs is accepted. In time slots for GPRS use. As one slot is assumed to provide
(2) Rtot is simply taken to be the total bit rate of the cell. constant throughput of 10 kbps, each cell offers a throughput of
60 kbps (without EGPRS). A probability for a user to connect
B. Traffic and User Satisfaction to a particular cell is equal for all cells and is noted with p.
Since all users are connected to some cell, p is obtained with
Number of terminals is given as an input to the model. Each
terminal in the network is using either streaming media,
1
browsing or messaging service. Streaming media service is one p= - . (6)
where user chooses typically a video clip from a server in a Ne
network. The clip is sent to the user, and viewed during the F or each cell the number of streaming users follows the
transmission. In this study the media encoding rate is assumed binomial distribution
to be constant 20 kbps. Browsing service is assumed to be web
surfing type of service. User is downloading content from the
Internet, but due to snapshot nature (passing of time is not (7)
modeled) of the model there is no need to take into account file
or reading time length. Browsing service is simply assumed to
require a certain throughput in order to not to make user The number of streaming users in a cell is independent of
annoyed by delays in page download. Finally messaging number of browsing or messaging users. Thus the probability
service is one where users send messages or e-mails where of having ns streaming, nB browsing and nM messaging users in
time is not a critical factor. a cell is given by

The three services have different required throughput in


order to achieve satisfactory QoE for the user. For streaming
service the required throughput is 20 kbps based on the media
encoding rate. It is also the maximum throughput taken by one
streaming TBF no matter how much capacity is available. For
browsing service the definition of minimum throughput giving
acceptable QoE is difficult to define. Thus several values are
used. They are listed in section IV. Messaging service is where Ns, NB and N M are total number of streaming, browsing
assumed to be such where throughput does not significantly and messaging users in the network and p is the probability of a
impact the QoE. Therefore the minimum throughput is set to 2 call to be located in one particular cell as defined in (6).
kbps, as this is assumed to be sufficient for example to avoid
problems with retransmissions on higher layer protocols, such Once the probability of each call combination for a cell is
as Tep (transmission control protocol). Maximum throughput known, the expected value of happy users within each service
with both browsing and messaging service is set to 30 kbps. class in the cell can be calculated. As the cells are identical, the
Mobile terminals are assumed to be capable to use maximum expected number of satisfied users in the network is obtained
three time slots in downlink direction, and one slot is assumed simply by multiplying with number of cells. Thus
to provide the throughput of 10 kbps (without EGPRS).
On network level it is required that within each of the three
hi = Ne L P (nS,nB,nM )n; , (9)

service classes, at least 90% of the users have to meet the


{(nS.nB.nM llR,2R,min}
throughput criterion. Otherwise it is considered that the
network level quality is unacceptable and there were too many
where h i, iE {S,B,M} refers to expected number of happy
users in the network. users within any service class, Ne is the number of cells in the
network, P is as defined in (8), Ri denotes the throughput of
Traffic mix in the network refers to portions of streaming,
service class i calculated with (3), R;min is the chosen minimum
browsing and messaging users in the network, denoted with VS,
VB and VM respectively. Multiple traffic mixes are tested to see throughput for service class i that provides adequate QoE and
if they impact the optimal values for SSS parameters of each ni refers to either ns, nB or nM-
service class. Given the total number of users in the network
Portion of satisfied users Si within service class i is given by
Niol and the traffic mix, the total number of users in the
network within each service class is obtained with
iE {S,B,M}. (10)
Ni = round (V;Ntot ), iE {S,B,M}, ( 4)

where round means taking closest integer value, S refers to It is required that
streaming, B to browsing and M to messaging users. Since Vi > min
represent portions of all users, it holds that Si -Si .
= 090 Vi (11)

0-7803-8521-7/04/$20.00 © 2004 IEEE 3157


for all three service classes, streaming, browsing and Traffic mix with x % margin refers to all traffic mIxes
messaging. It is assumed that any service having too low user where such points are removed where
satisfaction rate is not widely accepted by users. Thus the
requirement for high enough satisfaction rate is independent for Vi < x/lOO ( 12)
each service class.
for any iE {S,B,M}.
To further clarify the model, all the input parameters are
listed in Table 1. With a given set of input values, the output of The grid in the space sPamled by Vs, VB and VM can be
the model is either 'positive' or 'negative', depending on represented in a triangle form as shown in Fig. 2. Each service
whether condition specified by (11) is true or not, type has its own vertex where the 100% of the traffic belongs
correspondingly. Positive result indicates that the network level to that particular service type. Starting from the vertex, the
quality was within acceptable limits. If negative result is portion of the traffic type in the traffic mix is decreasing along
obtained, Ntot is decreased and the model recalculated, until the line from the vertex to the mid point of the opposing edge.
positive result is received. Maximum value of Ntot that still Fig. 2 shows the optimalSSSB values for all traffic mixes with
gives positive result is called the maximum load of the network 5% R;in fixed to 10 kbps. Additionally in the results
grid and
with all other inputs fixed, Nmox. Maximum load can be found shown by Fig. 2 SSSB is constrained to have identical value in
out for all scheduling strategies, i.e. all feasible values of SSSB all the cells in the network.
(based on work done in [4] it is evident that SSSs can be
always fixed to best priority, 1, and SSSM to worst priority, 12). From Fig. 2 it can be seen that highest values (worst
Since there are only 12 feasible values for SSSB, brute force priority) for SSSB are on the area where there are about to 25%
method is used to find out the value giving highest maximum 40% of streaming traffic and more browsing users than
load. SSSB can be constrained to have identical value in all the messaging. Low SSSB is optimal when there is only little
cells in the network. It can be further constrained to have streaming or browsing traffic is present. This indicates that
identical value over all traffic mixes. When such constraints are browsing traffic can have relatively good priority when there is
used, they are noted in connection with results. The value of no real competition between streaming and browsing service
SSSB that maximizes Nmox is called optimal SSSB. types due to small presence of either type. When there are
significant portion of both traffic types, then browsing priority
has to yield to make room for streaming.
IV. RESULTS

Nmox is obtained for each traffic mix. Value of Nmox depends


In this study the optimal SSSB is searched for all points on a
grid in the space spanned by Vs, VB and VM- Distance between on constraints set on SSSB. Four different prioritization
strategies are studied in this paper. In the benchmark scenario
the grid points has been 0.05 (i.e. 5% step). Note that due to
there is no differentiation at all, so that each service receives
constraint given by (5), this space is two-dimensional only. The
equal treatment. We call static differentiation strategy the case
grid is referred to as 'all traffic mixes' further in the text.
where SSSs and SSSM are set according to Table I, and SSSB is
optimized to maximize Nmox but is constrained to single value
TABLE!. INPUT PARAMETERS FOR THE MODEL. in all cells and all traffic mixes. It corresponds to real network
Symbol Value Explanation situation where SSS values are used but they are on the BSC
level and cannot be set on cell-by-cell basis. Furthermore SSS
Rmin
s
20 kbps Min acceptable throughput for streaming values are kept fixed irrespective of the average traffic mix in
Rmax
s
20 kbps Max throughput used by streaming TBF the network.

Rmin
B
Varies Min acceptable throughput for browsing Dynamic differentiation strategy differs from static one in
Rmax
B
30 kbps Max throughput used by browsing TBF that SSSB is individually optimized for each traffic mix. Still it
is constrained to have an equal value in all cells of the network,
Rmin
M
2 kbps Min acceptable throughput for messaging
but this value is changed as traffic mix changes in the network.
Rmax
M
30 kbps Max throughput used by messaging TBF Finally in adaptive differentiation strategy, SSSB is optimized
Rtot 10 kbps Throughput provided by one cell individually for each cell. In real network it corresponds to a
case where SSS values are automatically optimized on cell
Ntot Varies Total number of users (TBFs)
level every time a new TBF is set up on the cell or an existing
Vs Varies Portion of streaming users (TBFs) one leaves.

VB Varies Portion of browsing users (TBFs)


TABLE II . SET O F FOUR TRAFFIC MIXES, NAMED 'EVOLUTION'.
VM Varies Portion of messaging users (TBFs)
SSSs 1 Scheduling step size for streaming TBFs Service Mix 1 Mix 2 Mix 3 Mix 4
SSSB Varies Scheduling step size for browsing TBFs Streaming 10% 15% 20% 25%
SSSM 12 Scheduling step size for messaging TBFs Browsing 40% 45% 50% 50%
p 11300 Probability of a call in any single cell Messaging 50% 40% 30% 25%
Smin 0.90 Minimum network level satisfaction level

0-7803-8521-7/04/$20.00 © 2004 IEEE 3158


Nmax is specific for each traffic mix. Therefore when
evaluating how much do aforementioned four prioritization
strategies make difference to Nmax , we have to consider some
particular traffic mix or rather a combination of several. Table
III presents the gain in Nmax in percents from using different
differentiation strategies compared to benchmark scenario.
Gain of 100% means that double traffic compared to
benchmark was carried with the same end user quality criteria.
The gains are listed for different values of Ry/in and for four
different traffic mix combinations. Traffic mix combinations
are defined by margin as explained in connection with (12),
except for traffic mix combination called 'Evolution'. It is a set
of four traffic mixes, listed in Table II. It represents a fictive
evolution path where portion of streaming and browsing traffic
types increase and portion of messaging decreases.
Figure 2. Optimal SSSB with Ry/in � 10 kbit/s
The results in Table III show that the load of the (E)GPRS
network can be increased significantly while maintaining
identical end user experience when service differentiation is V. CONCLUSIONS
introduced. The gains from static differentiation strategy vary
In this study we derived a model to estimate the quality of
from 80% to 211%, depending on Ry/in and on which traffic end user experience in (E)GPRS network with a given load and
mixes the results are averaged over. Gains from dynamic and traffic mix. Three distinct types of service were used;
adaptive strategies compared to benchmark case of not Streaming video, web browsing and messaging. A weighted
differentiation are very close to the gain from the static round robin (WRR) scheduling algorithm where priority of the
differentiation strategy. This means that the benefit from these connection depends on the service was presented. Three
more complex strategies is very marginal. The values are listed different strategies to set the WRR weights were compared.
with GPRS assumptions, but due to simplicity of the model the The derived model was used to evaluate the gain achieved with
results are similar with EGPRS assumptions (Rtot 30 kbps etc.). the three strategies in terms of carried load at a given quality of
end user experience. Benchmark scenario was the same
In Table III lower value of Ry/in always yields larger network with no prioritization, where each user had an equal
gains, since this allows browsing traffic to be treated with priority independent of the service used. The results showed
lower priority (higher SSS), thus leaving more resources for that the simplest (static) WRR strategy allowed roughly twice
streaming and background services. Increasing margin the load of the benchmark scenario to be carried in the network
decreases the gain with all differentiation strategies. This is due with equal end user experience. More complex WRR strategies
to fact that traffic mixes with very large portion of messaging showed only marginal further improvement compared to the
service can lead to extremely high gains. Increasing margin simple strategy.
leaves these extreme cases out thus decreasing the gain.
REFERENCES

TABLE III. G AIN FROM THREE QoS DIFFERENTIATION STRATEGIES. [1] Z. Jiang, Chang L. F., Shankaranarayanan N. K., "Providing multiple
service classes for bursty data traffi c in cellular networks", IEEE

Ry/in [kbps]
Static Dynamic Adaptive Infocom, 2000, pp. 1087-1096.
margin [%] SSSB [%] [%] [%] [2] W. Ajib and P. Godlewski, "Service disciplines performance for WWW
6 0 6 163 166 178 traffic in GPRS system," 3G Communication technologies, 2000, IEE
Conference publication No. 471, pp. 431-435.
6 5 6 137 138 145
6 10 8 112 113 114 [3] D. Todinca, P. Perry and J. Murphy, "Novel prioritised EGPRS medium
access regime for reduced file transfer delay during congested periods,"
6 'Evolution' 7 211 212 220 3G Communication technologies, 8-10 May 2002, IEE Conference
10 0 3 120 123 127 publication No. 489, pp. 550-554.
10 5 4 107 109 112 [4] A. Kuurne, R. Sanchez, D. Fernandez, "Service Based Prioritization in
10 10 4 95 96 99 (E)GPRS Radio Interface", IEEE VTC fall 2004 [in press].
10 'Evolution' 4 162 165 172 [5] C. Lindemann, A. Thiimmler, "Evaluating the GPRS radio interface for
12 0 3 98 101 108 different quality of service profiles", Proc. 12th GIIITG Fachtagung
12 5 3 88 89 96 kommunikation in verteilten systemen, Hamburg, Germany, Feb 2001,

12 10 3 80 81 87 pp. 291-30 l .

12 'Evolution' 3 121 121 134 [6] T. Halonen, J. Romero, J. Melero, (ed.) "GSM, GPRS and EDGE

- performance", John Wiley & Sons, 2002, ISBN 0470 84457 4.


8-12 0 3 115 117
8-12 5 3 102 104 -
8-12 10 4 91 93 -
8-12 'Evolution' 3 151 154 -

0-7803-8521-7/04/$20.00 © 2004 IEEE 3159

Вам также может понравиться