LTE Radio Load versus User Throughput

Jari Salo

A frequently asked question among LTE radio planners is that of how to determine the maximum acceptable
LTE radio interface load that should not be exceeded in order to maintain some targeted user data throughput.
This white paper summarizes some simple formulas for calculation of downlink user throughput as a function
Physical Resource Block (PRB) utilization. The formulas are expressed in terms of standardized 3GPP KPIs and are
hence computable from network performance counters. Examples from live LTE networks are given to illustrate
the usefulness.

1. INTRODUCTION throughput over LTE radio interface. In par-

A frequently asked question in radio planning
community is that of "what is the maximum LTE
radio load so that user throughput is still ac-
ceptable". A simple answer to the question is
available by exploiting results from IP network-
ing literature and basic queuing theory. In or-
der to make those results practically useful for
radio planners, the goal of this white paper is to
state the existing theoretical results in terms of
LTE radio utilization metrics that can be easily
computed from network performance statistics
available in commercial LTE base station prod-
ucts.
User throughput for elastic traffic transmit-
ted over fixed bandwidth non-wireless links
has been thoroughly investigated in engineer-
ing literature, an overview and references can be
found in [1]. Wireless network user throughput
has been analyzed already in [2] and in other
works of the authors thereof. The general idea
has been to assume that the underlying trans-
port protocol (usually TCP) realizes fair sharing
of the link bandwidth between users, and there-
fore the instantaneous user throughput is sim-
ply the total link bandwidth divided by the in-
stantaneous number of users. As the number of
active users sharing the link varies over time the
main challenge is how to find the average user
throughput over time.
In this paper the focus is on the average user
ticular the so-called M/G/1 Processor Sharing
(PS) approach is used to express user through-
put in terms of Physical Resource Block (PRB)
utilization. Importantly, the formulas are pre-
sented using the 3GPP standardized Key Perfor-
mance Indicators (KPI) and measurement exam-
ples from three live LTE networks are shown to
illustrate their utility. The results can be used for
both FDD and TDD variants of LTE. Although
in principle the general approach is applicable
to both downlink and uplink, the uplink case
is dependent on power control and link adapta-
tion, which leaves more degrees of freedom to
be considered. For this reason, throughout the
paper throughput refers to the downlink case,
with the uplink applicability left for the reader
to consider case-by-case.
No new science that cannot be found some-
where in the depths of engineering literature
is developed in this white paper. The target
audience are radio planning and optimization
engineers with a working knowledge of LTE.
Principles of LTE, as detailed in 3GPP specifica-
tions, will not be repeated in this paper. Instead
the reader is referred to the 3GPP standards or
the latest editions of well-known literature ref-
erences, e.g. [35] and many others.
In Section 2 the average user throughput is
formulated as a function of the number of active
data flows sharing radio scheduler resources. In
Section 3 these results are related to physical re-

1
2 J. Salo
source block (PRB) utilization by means of the
well-known M/G/1 processor sharing formula,
resulting in 11 degradation in user through-
put with denoting PRB utilization. Live net-
work examples are shown to validate the result.
The basic M/G/1 PS formula can be extended
to cover the case where voice and data traffic
are mixed. Other interesting use cases include
external rate limitation and carrier aggregation
which are discussed in Section 4. Further food
for thought is served in Section 5 where the va-
lidity of the underlying assumptions are scruti-
nized. Finally, conclusions are presented.
Version: 8 July 2016
Abbreviations
FDD Frequency Division Multiplexing
IP Internet Protocol
KPI Key Performance Indicator
LTE Long Term Evolution
M/G/1 Memoryless/Generic/single server
M/G/R Memoryless/Generic/R servers
OSS Operations Support System
PRB Physical Resource Block
PS Processor Sharing
RRC Radio Resource Control
TCP Transmission Control Protocol
TDD Time Division Multiplexing
TTI Transmission Time Interval
UE User Equipment
2. THROUGHPUT VERSUS NUMBER OF
ACTIVE USERS
In this section two different single-user
throughput metrics are defined. It is shown via
a live network example that the so-called sched-
uled IP throughput, standardized by 3GPP, can
be accurately predicted based on the number of
active UEs.
2.1. Basic Assumptions
The following assumptions are made.
Full buffer traffic model so that, in the ab-
sence of other users, a single user obtains
all available radio resources. For example,
constant bit rate streaming traffic would
not satisfy this condition.
If there is more than one user, the sched-
uler shares radio resources equally, on av-
erage, between users. This fair sharing
principle is assumed independently of the
radio conditions of the UEs sharing the
scheduler resources.
The first assumption will be relaxed in Sec-
tion 4, and the second assumption will be dis-
cussed in Section 5. The notion of "radio re-
source" could be defined in various ways, but in
case of the LTE radio interface it is convenient to
choose Physical Resource Blocks (PRBs) as the
resource being shared. For LTE downlink, PRB
utilization can be equated with transmit power
utilization as long as physical layer and com-
mon channel overhead are properly taken into
account.
With the assumptions above, instantaneous
user throughput with x active users download-
ing simultaneously would be 1x of the maximum
throughput. A UE is said to be active if there
are data remaining in the transmit buffer1 . As
different UEs in the cell start and finish their
data transfers, the number of active UEs (x), and
hence also the instantaneous user throughput,
changes over time. The interesting metric is the
average throughput experienced by UEs in the
cell.
2.2. Two user throughput metrics
Consider a UE located somewhere in a cell
experiencing certain radio quality. In the ab-
sence of any other users the UE is allocated all
available PRBs and receives some throughput
T1 , where the subscript 1 emphasizes that there
is one active user, i.e., x = 1. If there were x > 1
active UEs in the cell, the throughput of the user
1 TermsUE and user are used interchangeably. The num-
ber of active UEs is different from the number of RRC-
connected UEs since a UE can be RRC-connected without
having any data to receive or send.
LTE Radio Load versus User Throughput (DRAFT)
would be T1 /x instead. The maximum achiev-
able user throughput, T1 , depends on the user
location in the cell, interference from other cells,
number of transmit antennas, and so on. The
average user throughput Tue is defined as the ex-
pected value of T1 /x for positive integer x, in
other words
[ ]
T
Tue = E 1 , x 1 , (1)
x
where E[] denotes expected value and T1 and
x 1 are the random variables being averaged.
The user radio conditions, and hence the max-
imum throughput T1 , can be assumed statisti-
cally independent of the number of active UEs
in the cell, and (1) can thus be written as
[ ]
1 Fig. 1 shows an example of two cells serv-
Tue = E[ T1 ] E , x 1. (2)
x ing a large number of smart-phone users. The
The term C = E[ T1 ] will be called cell capacity
in this paper.
Unfortunately, the second term E[1/x ], which
is the average of the inverse of the number of
active UEs, cannot be always computed since it
is not usually available as a radio counter. On
the other hand, the average active UEs, E[ x ], is
a standardized KPI defined in 3GPP TS 36.314
and thus commonly implemented in commer-
cial systems. Therefore a more practical metric
results if E[1/x ] in (2) is replaced by 1/E[ x ], or
E[ T1 ]
Tsch = , x 1. (3)
E[ x ]
In 3GPP TS 36.314 this is called "scheduled
IP throughput". It should be emphasized that
E[ x ] = E[1/x ] and for this reason Tsch and Tue
are different throughput metrics and not equal
in value. The scheduled throughput can also be
written as [6]
S TION
Tsch = , (4)
W In the previous section, it was shown that av-
where S is the average file size (bytes) and W is
the average file transfer time. This form of the
scheduled throughput is often used in fixed-line
3
IP network throughput analysis where it goes
by the name "flow throughput". An application
to wireless network setting can be found [2] and
many others published since.
Regardless of which variant is used it is worth
noting that the scheduled throughput Tsch is al-
ways lower than user throughput Tue . This is a
direct result of the concavity of 1/x and Jensens
inequality: E[1/x ] > 1/E[ x ] and subsequently
Tue >Tsch . Typically one would be more inter-
ested in the end user experience making Tue
preferable to Tsch . Despite of this the focus in
this paper is on the scheduled throughput Tsch
because Tue is not usually computable from LTE
base station performance counters. A discussion
on the differences between different throughput
metrics can be found in [6,8].
horizontal axis is the average number of active
UEs, E[ x ], that has been extracted from hourly
OSS performance counters over a period of two
weeks. The hourly averages of x have been
binned to integers and for each bin the average
flow throughput is plotted. The flow through-
put shown on the vertical axis is also obtained
from the OSS counters and computed according
to the scheduled throughput (Tsch ) definition in
3GPP TS 36.314. Each dot in the figure is the av-
erage UE throughput for the horizontal binned
value of E[ x ]. It can be seen that the scheduled
throughput scales approximately inversely pro-
portional to the average active UEs, as predicted
by theory.
From here onwards, unless otherwise men-
tioned, the terms scheduled throughput, flow
throughput, user throughput, and UE through-
put are all used interchangeably to refer to Tsch .
3. THROUGHPUT VERSUS PRB UTILIZA-
erage scheduled throughput is inversely propor-
tional to the average number of active UEs. For
most radio planners the number of active UEs
4 J. Salo
Scheduled throughput versus number of active UEs
100
Scheduled throughput, Megabits per second
10
1 web page would generate scheduler PRB load
0.1
0 50
average number of active users, x
Figure 1. Scheduled throughput Tsch versus av-
erage number of active UEs, two examples from
live network.
is perhaps not the most intuitive KPI for charac-
terizing network load. A more commonly em-
ployed yardstick of network load is the fraction
of utilized PRBs. It would therefore be handy
for a radio planner to have some rule that re-
lates PRB utilization to user throughput. This is
the topic of the section.
3.1. Definition of Radio Load
In fixed link throughput analysis, the link uti-
lization is usually expressed in terms of av-
erage flow size S (bits) and flow arrival rate
(1/sec) as
S the number of data flow arrivals per time unit,
= ,
C
where C is the link bandwidth (bits per second).
The resource shared between users in this case
is the link bandwidth C. For LTE radio interface
another option is to use LTE physical resource
blocks (PRBs) instead. In this context, a 20MHz
LTE cell scheduler can be considered a "proces-
sor" that shares the PRBs equally between active
UEs. For a practitioner the benefit is that PRB
utilization is available from system counters in
any practical LTE system.
To map bytes to PRBs the file size S needs to
be converted to PRBs. For example, given an
end user spectral efficiency of = 1 bits per
second per Herz, one PRB can fit 180 bits of
end user data after channel coding and phys-
ical layer overhead. Hence a one MegaByte
20MHz cell
of 106 8/180 44000 PRBs. More generally,
10MHz cell
1 S
= , (6)
180 Rprb
100 150 200 250 300 350
where is the average cell spectral efficiency in
bits per second per Hz, Rprb is the PRB rate (e.g.,
105 PRBs per second for a 20MHz cell), and the
scaling factor 180 comes from the standardized
PRB bandwidth of 180kHz.
In daily usage it is not necessary to know the
traffic parameters and S since the PRB utiliza-
tion can be extracted from network statistics
directly. This is in contrast to cell capacity C
which is somewhat less straightforward to esti-
mate from counters.
3.2. Throughput versus PRB utilization via
M/G/1 PS model
To relate the number of active UEs, E[ x ] or
1/E[ x ], in (1)(3) to PRB utilization, some sta-
tistical assumptions on the arrival rate of the
user data flows need to be made. In the IP
engineering literature, it is common practice to
model TCP flow throughput in wireline links
using the so-called M/G/1 Processor Sharing
model. The M/G/1 PS model assumes that ,
(5) has Poisson distribution. This is typically as-
sumed well-justified, one reason being that it
leads to simple formulas. In the sequel, the
same approach is applied to LTE radio through-
put.
It is well-known from basic text books [7] that
under the assumption of Poisson distributed
the proportion of time with x active UEs is
x = (1 ) x , where x is a non-negative in-
teger and is the load defined in (6). With this
LTE Radio Load versus User Throughput (DRAFT)
Comparison of user throughput and scheduled throughput
1
Throughput as a fraction idle cell throughput C
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2
Figure 2. Comparison of the throughput metrics
in (7) and (8).
information the expected values E[1/x ] and E[ x ]
in (2) and (3) can be computed. The details can
be found in a number of references and the re-
sult for user throughput is [7,6,8,2]
C (1
Tue =
while the scheduled throughput is simply
Tsch = C (1 ) .
As mentioned, measurement of user through-
put Tue is not very straightforward and not typ-
ically implemented in commercial base station
products, while the scheduled throughput Tsch
is simpler to compute and has been standard-
ized by 3GPP [9]. Comparison of the two KPIs is
shown in Fig. 2. For a given , user throughput
is always higher than the scheduled throughput
as mentioned in Section 2.
3.3. Live network examples
Fig. 3 illustrates scheduled throughput Tsch
for six cells from three different networks. PRB
5
utilization is extracted from hourly counter
user throughput Tue measurements collected over a period of two
scheduled throughput T
sch
weeks and binned to 2PRB granularity. Each dot
presents average scheduled throughput of the
PRB bin computed based on the 3GPP method
[9]. It can be seen that the theoretical model (8)
fits measurements fairly well, and throughput
falls approximately linearly as a function of ra-
dio utilization. For example, at 50% utilization
user throughput has dropped to half of the cell
throughput while for 75% radio load a single
user receives on average only 25% of maximum
throughput. Such simple rule of thumbs can
0.4 0.6 0.8 1 provide useful capacity management guidance
PRB utilization,
for LTE networks, including traffic steering be-
tween different frequency layers.
In some cases cell-level measurement data do
not provide good fit to theory. Typically this
appears to be a result of unstable cell statistics,
for example due to short-term irregularities in
traffic or network interference. Such anomalies
violate some of the assumptions used to derive
the theoretical model. A noteworthy example,
to be explored further in the following section, is
where cell throughput is higher than the achiev-
able UE throughput. For example, if the cell
throughput C was 50Mbps it could happen that
) 1 the user can only achieve 50% PRB utilization
ln , (7)
1 and, say, 25Mbps throughput due to a band-
width limitation somewhere along the end-to-
end data path. The actual mechanism behind
(8) the user throughput throttling is not relevant
as such2 . A limitation of this kind violates the
full buffer assumption listed in Section 2, since
adding another active UE to the cell would not
halve the first users PRB resources. In the next
section the M/G/1 PS based formula (8) will
be extended to overcome this issue, albeit at the
cost of the lost simplicity.
2 Commonly encountered throttling mechanisms include
TCP window size limitation, limited bandwidth due to
capped subscription, transport bandwidth bottleneck, or
server-side throughput limitation.
6 J. Salo

Scheduled throughput versus PRB utilization

Megabits per second

15 30

10 20

5 10

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Megabits per second

30 20

15
20
10
10
5

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Megabits per second

10 10

5 5

0 0
0 20 40 60 80 100 0 20 40 60 80 100
PRB utilization, (percents) PRB utilization, (percents)

Figure 3. Scheduled throughput versus PRB utilization, six cells from three different networks.
LTE Radio Load versus User Throughput (DRAFT)
4. EXTENSIONS
Due its simplicity the M/G/1 PS formula
given in (8) is not directly applicable in some
practical use cases of interest. These cases in-
clude:
User throughput is not limited solely by
the available PRBs because of some ex-
ternal throttling mechanism such as maxi-
mum subscribed rate or bandwidth bottle-
neck in the end-to-end packet path.
Prioritization between users in radio
scheduler, for example mixing of voice and
best effort data traffic.
Carrier aggregation
In this section some extensions of (8) to the
above-mentioned cases are outlined. An ap-
plication to traffic balancing between frequency
layers is also briefly considered at the end of the
section.
4.1. External user throughput limitation
Consider the case where, due to some
throughput limiting mechanism, a UE is able to
use only 1/Rth portion of the cell PRB resources
even when there are no other active UEs. A suit-
able model in this case is the M/G/R PS, where
R defines the fraction of PRBs a UE is able to uti-
lize. Again skipping details, the M/G/R proces-
sor sharing version of the scheduled throughput
in (8) becomes [10]
C load from other users, for example for R = 5
Tsch = E2 ( R,R)
, (9)
1+ R (1 )
where E2 ( R, y) is the famous Erlangs second
formula
yR R
R! Ry
E2 ( R, y) = . (10)
R 1 R i yR
i=0 i! + R! RRy
Here y is a positive real number, R = 1, 2 . . . is a
positive integer and notation y = R was used
7
for brevity. Setting R = 1 results in the special
case of (8).
It can be observed that the seemingly innocent
constraint of external throughput limitation re-
sults in a considerably more involved formula
that can no longer be calculated using pencil-
and-paper. Another unpleasant finding is that R
is forced to be an integer which forces the user
PRB limit 1/R to be 12 , 31 , . . . which is unneces-
sarily coarse for practical use. Fortunately, it is
possible to generalize (10) to real-valued R. For
example [11] gives the formula
RK ( R, y)
E2 ( R, y) = , (11)
R y [1 K ( R, y)]
where
g( R + 1, y)
K ( R, y) = , (12)
1 G ( R + 1, y)
with g(, x ) and G (, x ) denoting the proba-
bility density function and cumulative density
function of the gamma distribution, respectively.
In (12) R 1 is a real number. Commonly
available spreadsheet programs implement the
gamma distribution functions which makes it
possible to compute values for (9) without spe-
cial mathematics software.
Figure 4 illustrates the formula (9). In the
figure, the user throughput on the vertical axis
has been normalized with the cell throughput
C = E[ T1 ]. It can be seen that with increasing
R user throughput becomes increasingly limited
by the external constraint and less impacted by
user throughput is one fifth of the cell through-
put until it starts to decrease at around 0.5.
Figure 5 shows an example of a cell with
high spectral efficiency and cell throughput in
the order of 50Mbps. As seen, the average
user throughput measured from counters sat-
urates at low , and consequently the simple
linear throughput degradation of (8) does not
hold well and at low load it is rather an upper
bound to the measured user throughput. Theo-
retical throughput is shown with the basic for-
8 J. Salo

Scheduled throughput using M/G/R PS for different R Scheduled throughput versus PRB utilization, M/G/R PS
1 50

Scheduled throughput, Megabits per second

0.9 45 R=1
Throughput (normalized with C)

0.8 40

0.7 35
R = 1, 1.2, 1.5, 2, 3, 5
0.6 30

0.5 25

0.4 20

0.3 15
R = 1.7
0.2 10

0.1 5

0 0
0 20 40 60 80 100 0 20 40 60 80 100
PRB utilization, (percents) PRB utilization, (percents)

Figure 4. Normalized scheduled throughput Figure 5. Scheduled throughput versus PRB uti-
versus PRB utilization, M/G/R PS model. . lization, M/G/R PS model. .

mula (R = 1) as well as the M/G/R based for- traffic. The approximation arises from the fact
mula (9) for R = 1.7 and C = 47 Mbps. It can be that the number of PRBs allocated for the GBR
seen that the M/G/R PS formula offers better traffic is time-varying (but approximated with
match to network-measured throughput values. its average value) since it depends on the num-
ber of GBR users and their radio quality. The
4.2. Priority-based scheduling of non-GBR quasi-static approximation is still usable if the
data and GBR traffic average GBR PRB utilization is less than, say,
A mix of Guaranteed Bit Rate (GBR) traf- 50% and varies slower than non-GBR utilization
fic and non-GBR elastic data traffic is a practi- [12]. This is a viable assumption since voice call
cal use case due to voice and video over LTE. holding time is usually in the order of 60 sec-
The main question is how the non-GBR data onds or more while non-GBR data transfers of
throughput changes due to GBR traffic that is smart phones last considerably shorter time.
scheduled with strict priority over the data (pos- As a summary, when GBR traffic is scheduled
sibly up to some capacity limit). Although a with strict priority over non-GBR traffic an ap-
considerable amount of research has been con- proximate way to calculate scheduled through-
ducted, no exact formulas are available for the put is to replace the PRB utilization in (8) with
case of two or more traffic classes scheduled the effective non-GBR PRB utilization eff , which
with strict priorities. Bounds and approxima- can be calculated as utilized non-GBR PRBs di-
tions can be found in e.g. [12,13] and refer- vided by the number of non-GBR PRBs available
ences therein. For daily radio planning work, after GBR scheduling.
the only usable approach seems to be the so-
called quasi-stationary approximation that as- 4.3. Carrier aggregation
sumes that the capacity available to non-GBR For carrier aggregation, the total user
traffic is simply the difference of the total num- throughput is the sum of component carrier
ber of PRBs and average PRBs allocated for GBR throughputs. For the simplest case of two car-
LTE Radio Load versus User Throughput (DRAFT) 9

riers, this leads to leaves the portion of 1 for the second cell.
For a given average file size S, the average sector
user throughput (8) can be written as a weighted
Tsch,ca = Tsch,1 + Tsch,2 sum
= C1 (1 1 ) + C2 (1 2 ) , (13)

where Ci and i are the average link throughput Tsec = Tsch,1 + (1 ) Tsch,2 (15)
and load of the ith link, respectively. This is eas- S S
ily generalized to multiple carriers. From (13) = + (1 ) (16)
W1 W2
it can be seen that the contribution of a highly = C1 (1 1 ) + (1 )C2 (1 2 ) , (17)
loaded carrier on total throughput is very small.
The cell throughputs Ci are not, strictly speak- where Ci and i are the average cell throughput
ing, independent since due to higher (lower) and load of the ith cell, respectively. The traffic
path loss a UE on a higher (lower) carrier fre- splitting ratio is assumed to apply to all traffic
quency will in general be closer to cell edge. and thus the cell loads are3
Thus in a given UE location C2 depends on
C1 experienced in the same location. Assum-
S
ing that the dependency is linear, C2 = kC1 , 1 = , (18)
for some constant k, the carrier aggregation C1
throughput gain from adding the second carrier S
2 = (1 ) . (19)
would be C2

All told, the average user throughput across the

Tsch,1 + Tsch,2
Gca =
Tsch,1
1 2
= 1+k . (14)
1 1
The constant k depends on cell geometry, path
loss difference, and other factors. Detailed in-
vestigation is beyond the scope of the present
discussion.
4.4. Traffic balancing between frequency lay-
ers
Considering two cells on different carrier fre-
quencies that cover the same physical area (e.g.
a sector), an interesting question is that of how
the cell loads are related to the average user
throughput of the sector. From earlier discus-
sion, the throughput in case of a single cell is
S Fig. 7. In the figure the total offered traffic
defined as Tsch = W where S is the average file
size and W is the average time to transmit the
file to the UE. To average the user throughput
over two cells, a traffic splitting ratio can be in-
troduced. The fraction of arriving data flows as-
signed to the first cell is denoted with which
two cells of the sector becomes
( )
S
Tsec = C1 1
C1
( )
S
+ (1 )C2 1 (1 ) . (20)
C2
Fig. 6 illustrates the result. Offered traffic is
10Mbps and C1 = 20Mbps. The average user
throughput is shown for different traffic split be-
tween layers, for a few selected values of C2 . It
can be seen that an optimum traffic split maxi-
mizing user throughput exists. Interestingly, for
C2 = 40Mbps all traffic should be carried by the
second layer in order to maximize average user
throughput.
The sector user throughput is also shown in
S = 10Mbps and the capacity of the first cell
C1 = 20Mbps are fixed, while C2 and traffic
3 Inthis section it is more convenient to express load in terms
of traffic volume and cell traffic capacity, rather than PRB
utilization.
10 J. Salo

traffic split versus throughput, C1=20Mbps, S = 10Mbps Optimum traffic balancing factor for two cells
30 0.5
sector average user throughput, Mbps

25
0.4

optimum traffic splitting factor

1 = 0.05, 0.25, 0.5, 0.75, 0.95, 1.0
20
0.3
15

0.2
10

5 C2 = 10Mbps, 20Mbps, 30Mbps, 40Mbps 0.1

0 0
0 0.2 0.4 0.6 0.8 1 1 1.5 2 2.5 3
traffic balancing ratio, ratio of cell capacities C /C
2 1

Figure 6. Average sector user throughput as a Figure 8. Optimum traffic splitting factor for
function of traffic split. C1 = 20Mbps, sector two parallel cells.
offered traffic is 10Mbps.

the even split, opt = 21 is optimum, as expected.

split are allowed to vary. The contours map Fig. 8 illustrates the optimum for differ-
parameter pairs (C2 , ) that result in the fixed
ent values of C C and 1 . It can be seen that
2
sector throughput shown in the contour label. 1

For example, when C2 = 20Mbps and = 0.5 if the capacity ratio C 2

C1 is higher than a certain
the sector average user throughput is 15Mbps. threshold, that depends on total sector traffic via
This should be compared to the case of one 1 , no positive opt exists and to maximize user
20Mbps cell in which case the user throughput throughput the first cell should not carry any
is 10Mbps as can be seen by looking at points traffic at all.
(0, 1) and (20, 0) in Fig. 7.
Figs. 6 and 7 invite the following question: 5. DISCUSSION
if the sector offered traffic S and the cell ca- In this section the following points are briefly
pacities are fixed, what is the optimum traf- discussed.
fic balancing factor that maximizes the sector
user throughput Tsec in (20) ? Skipping some Impact of network load and impact of in-
straightforward calculations the optimum split- terference from neighbour cells
ting ratio turns out be
The assumption of equal PRB sharing be-
C2 tween active UEs
1 1 C1
opt = + , (21)
2 4 1 The assumptions underlying the M/G/1
PS model
where 1 = S
C1 , i.e., the load of the first cell if the
second cell was non-existent. The optimum traf- 5.1. Impact of Network Load
fic split depends only on the ratio of cell capac- In (8) the and Tsch denote the PRB utilization
ities, not on their actual values. When C1 = C2 , and user throughput of the serving cell. Noth-
LTE Radio Load versus User Throughput (DRAFT) 11

Average user throughput in the sector. C1 = 20Mbps, S=10Mbps

1

0.9
proportion of traffic carried by the first cell,

15
0.8
15
10

0.7 20

0.6

0.5
25
15

20

0.4
10
5

0.3

0.2
25

0.1
15

20
10
0

0
0 5 10 15 20 25 30 35 40
capacity of the second cell C , Megabits per second
2

Figure 7. Equal throughput contours for two cells for different traffic balancing ratios. The first cell
capacity is constant C1 = 20Mbps, and the total offered traffic is 10Mbps. The solid contours indicate
(C2 , ) parameter pairs with constant user throughput across the two cells (Mbps). The red dashed
line indicates optimum traffic split opt that maximizes (20).
12
ing is said about the load of the surrounding
cells. This raises the question of how Tsch =
C (1 ) behaves when the neighbour cell load
also changes at the same time with . Interfer-
ence received from neighbour cells degrades the
spectral efficiency in the serving cell, hence it
is expected that the cell capacity C decreases as
neighbour cell load increases. Decreasing also
increases the PRB utilization (6) since the num-
ber of bits per PRB decreases (for a fixed traffic
volume S in bytes) . However, the change in
due to interference is indistinguishable from a
change due to traffic volume, and therefore in
this sense it does not directly impact through-
put calculation. The second order impact still
remains since increasing own cell PRB utiliza-
tion generates more interference to neighbours,
which in turn again increases own cell PRB uti-
lization.
Increased load can also have positive impact
on spectral efficiency, namely in the form of
multi-user diversity gain, where the scheduler
exploits frequency selectivity of the wideband
radio channel. UEs are opportunistically sched-
uled on the frequency subband that has the
highest relative channel gain for that UE. The
multiuser diversity increases with the number
of UEs scheduled per TTI, which was related to
PRB utilization in Section 3. Cell capacity gains
of up to 50% have been simulated [14]. The im-
pact on the present discussion is that if the serv-
ing cell load increases at the same time with the
neighbour cell load, degradation in spectral ef-
ficiency is partially compensated by multiuser
scheduling gain.
Depending on implementation the radio
scheduler may also trade off spectral efficiency
to latency by using free PRBs to transmit with
lower channel coding rate. This improves la-
tency since the probability of retransmission is
reduced, but it also reduces spectral efficiency at
low load. On the other hand, at high load most
PRBs tend to be in use and the scheduler is thus
forced to operate at higher spectral efficiency.
An example of the dependency of spectral ef-
J. Salo
PRB utilization versus spectral efficiency
2.5
cluster average spectral efficiency, bps/Herz
2 2.6GHz
800MHz
1.5
1
0.5
0
0 0.2 0.4 0.6 0.8 1
cluster average PRB utilization
Figure 9. Load versus spectral efficiency at clus-
ter level, inter-site distance 500 meters.
ficiency on load is shown in Fig. 9. Each dot
in the figure presents one hour average of ver-
sus for a cluster of 16 three-sector sites, with
two cells per sector, on 800MHz and 2.6GHz.
It can be seen that going from unloaded to 50%
load the decrease in spectral efficiency is around
30-40%. Characteristically for cellular networks,
the load in the plotted cluster is unevenly dis-
tributed and many cells have load near 100%,
while for others < 20%. Therefore a 50% clus-
ter average load can be considered quite high.
All in all, integrating the impact of neigh-
bour cell load in the throughput formulas seems
intractable. Measurement data from smart-
phone dominant networks and mass events in-
dicate that the scheduled throughput degrades
roughly linearly with the serving cell PRB uti-
lization, which implies that the linear degrada-
tion predicted by C (1 ) is still a useful ap-
proximation for most network operations pur-
poses.
LTE Radio Load versus User Throughput (DRAFT)
5.2. Assumption of Equal Sharing of PRBs
One of the assumptions listed in Section 2 is
that the scheduler shares PRBs equally between
active UEs in time and frequency domain. If this
assumption does not hold the user throughput
will not scale as 1/x where x is the number of
active UEs. For example, a scheduler that priori-
tizes UEs having high good signal quality would
not satisfy this requirement. Another commonly
implemented scheduler type, the Proportionally
Fair scheduler has been shown to allocate, on
average, resources equally between users in [15].
5.3. M/G/1 PS modeling assumptions
As with all models to obtain any usable re-
sults some assumptions need to be made. In
the case of M/G/1 PS the assumption is that
number of active UE arrivals per second is Pois-
son distributed, which effectively means that
as approaches 100%, also x increases sharply
which in turn makes user throughput approach
zero. In fact if the Poisson assumption holds ex-
actly, the average active UEs is given by [7]
E[ x ] = , x 0.
1
This could be considered a reasonable assump-
tion, especially in mass events and cells dom-
inated by smart phones. The other extreme is
a network (or a cell) where USB data modems
or wireless routers are the dominant UE type.
A single active UE downloading a very large
file, or a wireless router shared by multiple
subscribers, can cause PRB utilization of 100%
for long periods of time, and at the same time
hourly radio counters will show high scheduled
throughput per user.
Empirical evidence from several networks
suggests that cells that do not accord to C (1 )
frequently still present linear degradation in ,
but at a less steep rate. A typical example is
shown in Fig. 10. The left-hand side panel
shows scheduled throughput and the M/G/1
prediction, and clearly the actual throughput
from counters has less steep decline with ;
the best linear fit results in 1 degradation,
13
where = 0.84. The mismatch is caused by
a small number of active UEs generating high
load, in other words while 1, x does not
tend to a high value (theoretically ), but in-
stead saturates at small value.
The right-hand panel of Fig. 10 illustrates the
more general formula (3) which does not rely on
the M/G/1 PS model and the underlying Pois-
son assumption. The throughput is simply the
cell capacity divided by the average of x, i.e., the
number of active UEs. The agreement with the-
ory is better than with M/G/1 PS, since no as-
sumptions are made regarding the distribution
of x.
To conclude this section, the extremely handy
and insightful linear degradation of scheduled
throughput (8) does not hold in cases where the
underlying statistical assumptions are violated.
The more general formula (3) makes no assump-
tions on distribution of the active UEs and is
therefore more robust.
6. CONCLUSION
This white paper discussed mapping of LTE
user throughput to radio utilization. The aver-
age scheduled throughput, calculated with the
3GPP method, was shown to be fairly accu-
rately predicted by the cell capacity divided by
the average number of active UEs. Adding the
usual assumption that user data flow arrivals
are Poisson distributed this result was expressed
in terms of Physical Resource Block utilization,
the outcome being the well-known M/G/1 Pro-
cessor Sharing formula that predicts linear de-
crease of user throughput with PRB utilization.
Real-world measurement examples were inter-
spersed throughout the paper to illustrate the
applicability of the results. Extensions to some
common LTE use cases were presented, and ap-
plication to load balancing between frequency
layers was discussed. A discussion on the un-
derlying assumptions concluded the paper.
14 J. Salo

REFERENCES

1. J. W. Roberts, A survey on statistical band-

width sharing, Computer Networks, 2004
2. T. Bonald and A. Proutiere, Wireless Down-
link Data Channels: User Performance and
Cell Dimensioning, Proc. ACM Mobicom,
2003
3. H. Holma et al (eds.), LTE for UMTS, Wiley,
2009.
4. F. Khan, LTE for 4G Mobile Broadband,
throughput versus PRB utilization throughput versus avg active UEs
Cambridge University Press, 2009.
30 100 5. S. Sesia et al (eds.), LTE, The UMTS Long
sch tput sch tput
25(1) 65/x Term Evolution: From Theory to Practice,
23.3(10.84)
25 Wiley, 2009.
scheduled throughput Tsch, Mbps

scheduled throughput Tsch, Mbps

6. N. Chen and S. Jordan, Throughput in

20 Processor-Sharing Queues, IEEE Tr. Aut.
Contrl, Feb. 2007
15 10 7. L. Kleinrock, Queueing Systems Vol. 1,2,
Wiley-Interscience, 1975
10 8. A. A. Kherani and A. Kumar, Stochastic
Models for Throughput Analysis of Ran-
5
domly Arriving Elastic Flows in the Internet,
Proc. IEEE Infocom, 2002
0 1
0 0.5 1 0 20 40 60 9. 3GPP TS 36.314, v11.1.0, Dec. 2012
PRB utilization () average active UEs (x)
10. K. Lindberger, Balancing Quality of Service,
Pricing and Utilisation in Multiservice Net-
works with Stream and Elastic Traffic, ITC
16, 1999
Figure 10. Example of a 20MHz cell where 11. V. Naumov and O. Martikainen, Queueing
M/G/1 PS formula does not hold. Left-hand Systems with Fractional Number of Servers,
side shows M/G/1 PS formula (8) with red Discussion Papers No 1268, The Research In-
dashed line, and best linear fit black dotted line. stitute of the Finnish Economy, 2012
Right-hand side shows (3) and C/E[ x ] fit with 12. T. Bonald and A. Proutiere, On Performance
C = 65Mbps. Bounds for the Integration of Elastic and
Adaptive Streaming Flows, Proc. SIGMET-
RICS, 2004
13. R. Malhotra and J. L. van den Berg, Flow
level performance approximations for elas-
tic traffic integrated with prioritized stream
traffic, Proc. International Telecommunica-
tions Network Strategy and Planning Sym-
posium, 2006
14. A. Pokhariyal et al, Performance of Down-
link frequency Domain Packet Scheduling
for the UTRAN Long Term Evolution, IEEE
LTE Radio Load versus User Throughput (DRAFT) 15

PIMRC, 2006
15. J. M. Holzman, CDMA Forward Link Water-
filling Power Control, Proc. IEEE VTC, 2000

Jari Salo received the degrees of Master of

Science in Technology and Doctor of Science in
Technology from Helsinki University of Tech-
nology (TKK) in 2000 and 2006, respectively.
From 2006 to 2013, he was with European Com-
munications Engineering Ltd, where he worked
as a consultant for mobile wireless industry and
operators. He is currently with Nokia Net-
works where he works as a radio and access
transport planner. He has written or co-written
5 book chapters, 14 IEE/IEEE journal papers,
35 conference papers, and is the recipient of
the Neal Shepherd Memorial Best Propagation
Paper Award from the IEEE Vehicular Tech-
nology Society, for the best radio propagation
paper published in IEEE Transactions on Vehic-
ular Technology during 2007. He is also a co-
creator of the wide-band spatial channel model
adopted by 3GPP for LTE system-level simula-
tions. E-mail: jari.1.salo@nokia.com