You are on page 1of 10

SPECTROSCOPY OF THE PACKET RADIO INTERFACES

A CASE STUDY WITH GSM/GPRS


Jorma Kilpi
VTT Information Technology
P.O. Box 1202, FIN 02044 VTT, Finland
E-mail: Jorma.Kilpi@vtt.
ABSTRACT
In this paper we apply some ideas presented in [1] of network spectroscopy to the analysis of a packet
radio interface. We analyze the IP packet interarrival time distribution of upstream trafc, taking into
account also the packet sizes in the case where the IP packet trace is measured after the Time Division
Multiple Access (TDMA) based radio interface of General Packet Radio Service (GPRS). When the
Logical Link Control (LLC) layer works in unacknowledged mode and the Radio Link Control (RLC)
layer works in acknowledged mode, we show that using the ideas of spectroscopy it is possible to some
extent to detect the behaviour of the TDMA based radio interface of GPRS from the IP level packet data.
More precisely, it is possible to extract the number of Packet Data Channels (PDCHs) and Channel
Coding Scheme used for uplink trafc of a single mobile station (MS). Physical connections between
the MS and the Base Station Subsystem (BSS), called Temporary Block Flows (TBFs), can be extracted
also, and it is possible to indirectly detect some retransmissions of radio blocks within TBF. The general
method is not restricted to TDMA and should give good results also for other packet radio interfaces
like Code Division Multiple Access (CDMA) based packet radio interfaces or Carrier Sense Multiple
Access (CSMA) based WLANs.

1 Introduction

ful also for analyses of WCDMA or CSMA based packet


radio interfaces.

This study was motivated by the ideas of Broido & al.


presented in paper [1]. They dened network spectroscopy as a branch of Internet science that deals with
object identication on the basis of delay, period and
frequency spectra. Packet delay analysis is one of the
application areas of network spectroscopy. In this paper we shall apply the ideas of [1], especially the use of
the Radon transform, to the analysis of the TDMA-based
packet radio interface of GPRS.

As one concrete application of our approach one


can consider a system which continuously monitors upstream trafc inside a GPRS backbone network. The
user IP packet is tunneled inside the GPRS backbone. It
should be possible to trace backwards the mobile routing
of the tunnel packet, which carries the user IP packet as
payload, and identify the cell where the Mobile Station
(MS) was when the packet was sent. This requires only
combining information from the Mobility Management
(MM). Such a system could, for example, at given time
instances select, according to some criteria or just randomly, some sessions or upstream ows, search for the
signs or indicators of problems in the radio interface, or
interfaces just after it, and report or even alarm in some
cases.

As a case study we will analyze the IP packet interarrival time distribution taking into account also the packet
sizes in case when the IP packet trace was measured after
the GSM/GPRS radio interface. The LLC layer worked
in unacknowledged mode and the RLC layer worked in
acknowledged mode. In that case, the interarrival time
distribution contains information about the performance
of the radio interface. We combine the technical knowledge of how the corresponding radio interface and those
interfaces closely related to it work with the statistical
analysis of the data trace. In this way we get a better understanding of those factors that are important in analysis
and modelling of the radio interface. We show examples
from upstream trafc of GPRS sessions. We also present
the main ideas of the data analysis procedure along with
the case study. However, the general idea and method
is not in any sense limited to TDMA and should be use-

Hence, using only a few monitoring positions, it


should be possible to see some performance characteristics of all active uplink radio interfaces of the whole
network. It is also reasonable to expect that the performance problems of uplink and downlink are often correlated, even in the cases when the uplink and downlink
should work independently, as with GPRS. Of course,
problems of the radio interface are more likely due to external factors or the MS than to the network. But, for
example, if one cell continuously shows worse performance than others, this would eventually show up. As

frame structure used in TDMA with GPRS, see the technical specication GSM 05.01 from [5].

Time Slot (TS)

Basic unit

This case study is based on examples from the data from


a GPRS trafc measurement trace which was captured
from the monitoring port of a rewall router between the
GPRS backbone network and the Internet access point.
The measurement was made in May 2002. See [2] for
further details of the measurement and a general description of the data. See also [3] for a similar but much more
detailed GPRS trafc measurement and [4] for a discussion of some measurement based performance issues related with GPRS.
The study is done per session, or sessionwise. In the
data analysis, an Internet session over GPRS was dened
by the temporary IP address given to the user in the activation phase of an IP context. These temporary addresses
were delivered in a round robin manner and, since the intensity of session arrivals was rather low during the measurement time, we can be certain that all trafc from the
same temporary IP address came from the same GPRS
MS. Unlike in [3] we were not able to measure the Packet
Data Protocol (PDP) contexts since, due to technical and
privacy issues, this would have been too demanding with
the resources available. Moreover, since in our case the
PDP is always IP, we will talk about IP context instead
of PDP context.
The 3rd Generation Partnership Project (3GPP) is
responsible for the technical specications of GPRS,
downloadable at ETSIs homepage [5]. At the time of the
measurement only the Release 97 version of the specications can be expected to have been lled.
The emphasis is on the uplink direction, which is the
same as the upstream direction for an Internet connection. The phrase mobile-originated is often used in the
3GPP specications. The upstream data was thus measured after the radio interface, but not immediately after.

TDMA Frame

8 TSs

The basic idea of our data analysis procedure is not very


complicated. However, it requires some quite detailed
knowledge of the Um, Abis and Gb interfaces and for
this reason we recall and point out briey those properties of GPRS that are important to know in this context
and/or are needed later in this paper, so that readers not
very familiar with GPRS could more easily understand
our data analysis procedure in Section 2.2. Readers already very familiar with GPRS can just check the formulas (1), (2) and (3) of this section before going directly to
the section 2.2.
Table 1 recalls the denitions and durations of the time

Multiframe

52 TDMA Frames

240

Superframe

8 multiframes

1920

Table 1: The TDMA time frame structure.


Every 13th TDMA frame inside a multiframe is reserved for other purposes than data transfer; they are either used for synchronizing the MS with the base station,
or for a MS to monitoring for a better base station with
better signal to interference ratio. The time slots of the
TDMA frame available for GPRS use are called packet
data channels (PDCHs). The multislot functionality of
GPRS allows a MS to be allocated more than one PDCH
simultaneously. PDCHs are dynamically and temporarily allocated to a MS for the transmission or reception of
data.
Figure 1 shows a part of the transmission plane of
GSM/GPRS. The network elements of the transmission plane that a packet sees on the way up are called
the Mobile Station (MS), the Base Station Subsystem
(BSS), the Serving GPRS Support Node (SGSN) and
the Gateway GPRS Support Node (GGSN). The protocol layers are Subnetwork Dependent Convergence Protocol (SNDCP), Logical Link Control (LLC), Radio Link
Control (RLC), Media Access Control (MAC), and the
TDMA based GSM Radio Frequency (RF). As the name
indicates, one role of SNDCP is to make the underlying
protocols independent of the higher layer protocols: TCP
(or UDP) and IP are not part of GPRS protocol stack.
They are present in gure 1 since they are the applications that are used in the data that we have.
The transmission plane in gure 1 is not complete,
since the BSS consists of one Base Station Controller
(BSC) and several Base Transceiver Stations (BTS). The
interface between a BTS and the BSC is called Abis, see
gure 3.
At the MS, more precisely at the users GPRS handset,
an IP packet sent upstream is divided into SNDCP segments, LLC frames and RLC radio blocks. When there
should not be any danger of confusion, we will simply
talk about segments, frames and blocks. See gure 1 as
mnemonic. One block is transmitted by the RF layer as
four bursts of bits where one burst of bits is the physical
content of one time slot. One block is always transmitted within one PDCH and the delay of one block at the
. After the uplink
Um interface is
Um and Gb interfaces the IP packet is reassembled from
4 2  1 %
53 (0)(&#"
' % $
!

2.1 The Um, Abis and Gb interfaces of


GSM/GPRS





2 Case study

Duration
(ms)

 


Denition

  
 

such it could be a complementary tool for a radio interface protocol analyzer, a physical device, which has to be
close to the base station in order to capture and monitor
the trafc of the radio interface.

TCP, UDP

GGSN

Figure 1: Transmission plane of GPRS. The up- and


downlink directions work independently.

bursts, blocks, frames and possible segments. At the network side this reassembly is not done at a single device;
each block is reassembled from four bursts of bits at the
BTS and forwarded to the BSC. There, frames are reassembled from blocks and relayed to the SGSN, where
segments are reassembled from frames. A segment may
consist of one or more IP packets of the same MS. The
SNDCP can optimize channel efciency by multiplexing
several small packets into one segment. It also divides
too large packets into several segments with the aim that
one segment ts within one frame. If the segment consists of one packet only, then the packet is in practice reassembled already at the BSC, but the SGSN is the place
where the packet is completely reassembled after the Um
and Gb interfaces. The GGSN is the rst place where
any action due to the IP packet header is done.
During the time of our measurement there was at least
one GPRS mobile phone available in the Finnish market that could use 2 PDCHs in the uplink radio interface. Moreover, this particular model had been available
at least a few months. On the other hand, it is unlikely
that any phone could have used more than 2 PDCHs in
the uplink.
A physical connection between a MS and a BSS is
called a Temporary Block Flow (TBF). The TBF is a unidirectional concept and consists of the allocation of one
or more PDCHs and the number of blocks to be sent or
received. The allocation can be xed or dynamic; by dynamic allocation, the number of blocks to be transmitted
is not xed beforehand. The TBFs are the main object of
our study.
The high Bit Error Rate (BER) of radio interfaces makes it necessary to use error correcting codes.
Technical specications introduce four channel coding
schemes, but only two of them, called CS-1 and CS-2,
were implemented at the network side during the time of
our measurement. CS-1 has a better error correcting capability and thus a larger coverage, but less user payload
bits at each block, and thus it offers a lower bit rate to the
user. Channel codings are used at the Um interface.
Given a packet of size bytes, the exact number of
bits of the packet at one block depends on the channel

Size of one LLC frame in bits


Bits in one block

(1)

Here
and
is just the size
of the packet, the payload of an LLC frame, in bits. The
value 40 is explained next.
Even the use of error correcting codes does not guarantee that the blocks (and frames) are correctly received
over the Um interface. To improve the reliability of data
transfer, both the RLC and/or LLC layers may work in
acknowledged mode; blocks (frames) are sent within a
window and are periodically acknowledged using a selective acknowledgement method allowing retransmissions of erroneously received blocks (frames). When the
data used in this paper were measured, the RLC layer
worked in acknowledged mode but the LLC layer did
not. It means that no reordering mechanism of the LLC
layer Packet Data Units (PDUs) was provided. Only
transmission and format errors of PDUs were detected,
and duplicate Unconrmed Information (UI) frames that
carried PDUs were discarded. The LLC layer PDU is
the same thing as a SNDCP segment, hence typically the
same thing as one IP packet or, sometimes, two small
packets multiplexed into one segment.
The smallest size of the UI frame without payload is
5 octets, 40 bits, see 3GPP TS 04.64 from [5]. We have
used this value as framing bits in formula (1). Luckily,
the formula (1) and our analysis are not very sensitive
to the exact value of framing bits; an error of 2-3 octets
would not be even visible in gure 2.
Nblock (B,CS)
60
50
Nblock

SGSN

!
)1

BSS

FR

RF

Bursts

G ( E @B @8 5 3 0
HFDCA976421)('

%
& 1 #

MS

MAC FR

BSSGP



RF

LLC

BSSGP

MAC

RLC

GTP

 
"! 

Blocks

GTP

Frames

LLC
RLC

IP
SNDCP

  


SNDCP

Segments

Packets

IP

coding scheme, the number of framing bits of the LLC


layer, and whether the framed segment consists of one or
more packets. It may vary a little. Let
be a parameter
with value
in case of CS-1 and
in
case of CS-2. These are approximate numbers of user
bits within one block. We need to calculate the number
of blocks that one packet of size bytes most typically is
assumed to generate,
, and use the following
simple formula: (see also gure 2)

Gi

% (%
1

Gn

Gb

Um

1


Network Elements, Protocol Layers and Interfaces of the Transmission Plane

CS-1

40
30
20
10

CS-2

0
0

200 400 600 800 1000 1200 1400


Packet Size (B)

Figure 2: Approximate number of blocks given the


packet size.

Let the parameter


be the number of PDCHs
temporarily allocated for the TBF; in the upstream case
we assume it takes only the values 1 or 2. The packet
delay of the Um interface is then approximately

Base Station Subsystem (BSS)

Um

Block

The SGSN needs to know the location of the MS before any radio resource allocation can exist and data be
transmitted or received. The network functions that take
care of the locating of the MS are called Mobility Management (MM). The MM states are called idle, stand-by
and ready states. After IP context activation, and if no
data is transmitted, the MS is in stand-by state. When the
MS needs to transmit (or receive) data, it rst changes
from stand-by state to ready state, and only then some
time slots from the Um interface are reserved. After releasing the radio resources both in the uplink and downlink there is a timer that changes the MS back to the
stand-by state unless a new allocation is requested. In
[3] it was observed that the MS usually stayed in ready
state several seconds after the data transfer was ended.
This information is important in our study since it gives
us some realistic upper bound to the packet interarrival
delay in the case when we have to decide whether two
successive packets have been transmitted within the same
TBF or not.
During the time of the measurement, GSM voice calls
had strict priority over GPRS when the time slots from
the TDMA frame were allocated. If there was contention
of available GPRS resources between several MSs, trafc from different MSs was multiplexed into available
PDCHs. This resource sharing, if it has occured, has
probably been fair between users and not based on different QoS proles.
The BTS contains the GSM RF layer and the Channel Codec Unit (CCU) which, in the case of an uplink block, decodes the channel coding scheme. The
block is then transmitted over the Abis interface to the
BSC, where the Packet Control Unit (PCU) differentiates GPRS blocks from GSM voice blocks; look at Figure 3. The RLC/MAC protocol layer is implemented in
the PCU. It turned out to be important in this study that
the transfer rate of a single block at the Abis interface has
always been 16 kb/s and the PCU frame, which carries
one decoded block, is always of size 320 bits, making
the transfer delay 20 ms. PCU frames are also sent with
periods of 20 ms. In this way, all 12 blocks that can be
transmitted within one multiframe of one PDCH get their
own PCU frame, and the total time, 240 ms, remains the
same. These values do not depend on the number of user
bits at the block, i.e. on the channel coding scheme used.

456 b

2  1 %



 
! 

2
3

This is because one block is always transmitted within

320 b

BTS

PCU
BSC

Figure 3: The Abis interface between BTS and BSC.

one PDCH, and each of the PDCHs has its own 16 kb/s
channel at the Abis interface. Compare with formula (2).
The transfer protocol of the Gb interface of gure 1
can be Frame Relay (FR), ATM, or even FR over ATM.
FR offers the bit rates 64 kb/s, 128 kb/s, 256 kb/s, 384
kb/s, 512 kb/s and 1984 kb/s. We do not know a priori the upstream (or downstream) transfer rate in the Gb
interface. Since it depends on how many BTSs are connected to one BSC, it need not be the same for different
physical Gb interfaces. We will see that the upstream
transfer rate can sometimes be inferred from the sessionwise data. In general we assume that one LLC frame is
transmitted within one FR frame, see gure 3.

2.2

Data analysis procedure

Our measurement was thus done from a fast interface


which, however, was not placed right after the slow Um,
Abis and Gb interfaces, see gure 1. After the Gi interface and before the true Internet access there may be a
small operators service network, which in the case of our
measurement contained at least the rewall router and a
Network Address Translation (NAT) box.
The accuracy of the time stamps is not as crucial as
in [1]; if they are accurate enough in the fast interface,
then they are accurate also for the modied idea that we
will present. The drift of the time stamping clock was
afterwards estimated to be approximately 19 ppm, and
before the analysis all time stamps were corrected by
multiplying them with the factor (1-19 ppm). In addition to this, there were other drifts in the time stamps
that were constant for different sessions. The origin of
these other drifts is uncertain but, if not due to the measurement setup, they may be due to different SGSNs or
the distance between the BSS and the SGSN, which can
in Finland probably be hundreds of kilometers.
Figure 4 below describes our basic idea. In Figure 4
the time refers to the time at the SGSN when the IP
packet is completely reassembled and the time refers
to the measured time stamp at the fast interface. The
measured time stamp is also associated at the end of
the packet since the Berkeley Packet Filter [6] on which
tcpdump is based does packet time stamping at the end



(3)

RF




The formula we use to calculate the packet delay of


the Abis interface is

CCU

FR
Frame





 
! 

(2)

Gb
Abis
PCU
RLC
Frame MAC

 


in

A
4@






6 ( 0 % 3 2 0 $
75)' 41"

9
8

@   

 

  

( % $ "
)' &#!







as the size dependent or deterministic part, and

9
C

as the residual or random part

A
4

of the interarrival delay . Our interest is on the cases


where is small, otherwise
could be ignored. Then
is typically quantized to some discrete set of possible
values.
Examples of single sessions are most interesting if
they contain both a large number of upstream packets
and packets of many different sizes. Table 2 shows
global characteristics of a few chosen examples. Some
of them are used in gures hereafter.
Figure 6 shows the packet interarrival time histogram
of Example 2. The 20 millisecond quantization due to
the Abis interface is rather exact. While one packet has
always required at least one TBF, our interest is in long
TBFs where at least two packets have been transmitted
within one TBF. We restrict our main interest to values of
smaller than 480 ms, the duration of two multiframes,
since then we can be almost certain that
and
have been transmitted within the same TBF. Figure 7
shows the
-plane (4) of Example 2.
The analysis continues like in [1] by taking a suitable
Radon transform of the empirical probability
,
dened in (5), for the simultaneous detection of the transfer rate of the Gb interface and of the parameters
and
. (See formulae (1), (2) and (3)). The Radon
transform in our case is dened as



  


 

  




3 %



R
SH


% ! 
  

4

2   '
1

% 

G
I

P
Q

  

H
I 5
G

E
F

A more serious problem than distortion is the buffering that occurs in the packets path before and at the measurement interface. However, at the BTS, there should
not be any buffering of blocks: erroneous blocks need
not be stored, and correctly received blocks are relayed
immediately to the BSC. At the BSC, there must be some
buffering capability of frames and blocks since, due to
RLC layer retransmissions, the blocks may not come in
the right order to the BSC. Hence the BSC must also have
reordering capabilities of blocks which may also induce

Figure 5: Maximum distortions as a function of constant


bit rate.

 
D 

50

40

 
D 

20
30
Rate (Mb/s)

10

Bi > Bi+1

-10

 
' 





 %

Bi < Bi+1

9


Writing the packet interarrival delay as


gure 4 we interpret

(5)




Dmax (ms)

     

10

-5

made from the measured values


,
, of
upstream trafc of a single session. See gure 4 for the
explanation why is associated to
and gure 7 for
a data example of the
-plane. We will analyze the
empirical probabilities
%

Maximum distortion

(4)

@  (  


of the packet.
We dene
and
. One
of the basic assumptions is that the size dependent distortion dened by
is small. Distortion
can be written as
, which
shows that it depends on the sizes of the packets
and
. The maximum increase occurs, when
and
bytes. The maximum decrease occurs,
when
and
bytes. Figure 5 shows
this maximum distortion as a function of the bit rate. We
assume that the interface Gn between SGSN and GGSN
has to be at least of multiplexing hierachy E3, and the
conclusion is that the size dependent distortion should
be less than
ms. We assume now that
and
do not distinguish and any more in the notation.

Figure 4: The grey boxes refer to the length of the


packet at the slow Um interface.

Time




t* ti+1
i+1

di

     

t* ti
i

Bi+1

@   ( 


d*
i

Bi

d*s
i

d*r
i

some packet size dependent delays. The GGSN acts as


a gateway between mobile packet routing and xed IP
routing of the Internet. The SGSN and the GGSN are assumed to have buffering capabilities like a router. Buffering of packets at the measurement equipment before the
time stamp is attached to the packet may also be a problem, see for example [7].
Luckily, during the time when our data was measured,
the intensity of GPRS trafc was so low that the buffering
of upstream trafc was about minimal at the network devices. Also, the buffering at the measurement equipment
was about minimal, even though this item was not optimized when the measurement was planned and made. In
section 3 we will discuss about planning of the optimal
measurement scenarios in the case of GSM/GPRS.
The data that we will use for analysis consist of pairs

d* = d*s + d*r
i
i
i

Packet
Duration
Volume
Mean Bit
% of TCP
Count
(kB)
Rate (kb/s)
Packets
Example 1
1155
19 min. 1 sec.
111.5
0.8
94.8%
Example 2
2259
65 min. 3 sec.
290.7
0.6
9.5%
Example 3
3360
10 min. 0 sec.
270.1
3.7
99.1%
Example 4
2525
29 min. 7 sec.
156.2
0.7
92.4%
Example 5
5529
29 min. 2 sec.
215.6
1.0
99.2%
Example 6
2808
11 min. 5 sec.
152.9
1.8
98.6%
Table 2: Upstream trafc characteristics of some sessions.

E


 

H
G




 

Number of Packets

H
G

(6)

Here the sum is taken over all those values of such


that
. The heuristics of the minimum entropy is that the best explanation reduces the randomness
most. Figure 8 shows that, from all four realistic alternatives for the values of the parameters
and
,
the one corresponding to CS-2 and 2 PDCHs is the most
promising one. Indeed, gure 9 shows that with CS-2
and 2 PDCHs the minimum is achieved with 385 kb/s,
which is 384 kb/s plus some error due to inaccuracies of
our formulas, distortion and measurement setup. Note
also that the value 384 kb/s is not contradictory to the
assumption that the distortion can be neglected since the
delay of the Gb interface was not included when we argued that
due to gure 5.
G

1440

Four Cases

960

5.2
5
H(v)

480

576
1024
Packet Size (B)

Figure 7: The

40 200

 
D 

4.8
4.6

1500

CS-2 and 2 PDCHs


4.4

-plane of example 2.
0

The term of above stands for the residual delay. The


term
is our simplied model of the packet delay
of the Gb interface assuming that it has transfer rate
b/ms, i.e. kb/s. Note that the use of a formula like
assumes a ow of bits with constant bit rate , which in
general is far from reality. It works in our case since
the physical Gb interface was not congested due to other
simultaneous sessions at the time. Since we do not know
the effect of distortion accurately enough we round the
size dependent term

500

1000
1500
v (kb/s)

2000

Figure 8: From all 4 different alternatives, CS-2 with 2


PDCHs is the most promising one for Example 2.


 % ! 
  

4 2  1
53 '

% 

G

P
Q

  


G
5

E
F


2 1 


2 1 

1
H


%  
 


4
52

 %

 (%
 1

to whole milliseconds (ms).


Like in [1] we detect the value of the (deterministic)

Next we describe what could be done if the accuracy


of the time stamps would be high enough and the effect
of distortion could be either ignored or corrected. Assume now that we have already identied the values of
and
. Namely, had there been congestion at
the BSC or at the Gb interface, we could have used the
Radon transform more like in [1] and, instead of looking
at the transfer rate of Gb interface, look for the time
that one LLC frame requires at the Gb interface:

Interarrival Delay (ms)

Example 2

1920

Figure 6: Histogram of packet interarrival delay of Example 2.

 

1920

H
G

480
960
1440
Interarrival Delay (ms)

10

  
H
I 5
G

E
1

of

  

Multiframe

100

1000

by looking at the minimum of Shannons

parameter
entropy

Example 2, Bin Size 1 ms


Superframe

Example 2

Example 1, Bin Size 1 ms

Number of Packets

1000

H(v)

4.3
4.2
4.1
4

100

10

3.9

1
350

450

480
960
1440
Residual Delay (ms)

1920

is 385

Figure 11: The overall shape is similar with Examples 1


and 2.

which minimizes
. This entropyand choose
minimizing could then be interpreted as the equilibrium time of one LLC frame at the Gb interface. Recall
from gure 3 that, at the BSC, one LLC frame is put inside the FR frame for the Gb interface. If we know the
transfer rate of the physical Gb interface and the size of
the FR frame, then could be used to detect whether the
BSC or the Gb interface was congested or not. This is because we could compare to the ideal case with no congestion. But, as already mentioned, this would require
more accurate time stamps and more precise knowledge
of the possible distortion so that we could control the errors.
After removing
from we get the residual values
. Figures 10 and 11 show two data examples
of the distribution of the residual delay . Compare gure 10 with gure 6. The rst block of each multiframe
of one of the downlink PDCHs is reserved to signalling
by default. If no other downlink signalling is used, it
explains partially the gap between 240 ms and 480 ms.

and raises some natural questions: How do we explain


negative values? How do we explain horizontal positioning? We already know that the 20 ms vertical shift
between horizontal layers is due to the Abis interface.

Residual Delay (ms)

200
100
0
-100
-200

A
C

Figure 12: The residual

Residual Delay (ms)

10

A
4

1
480
960
1440
Residual Delay (ms)

-plane of Example 2.

Example 3, Residual (B,d)-Plane

100



9
4

200 400 600 800 100012001400


Packet Size (B)

Example 2, Bin Size 1 ms

1000
Number of Packets

Example 2, Residual (B,d)-Plane

  

that minimizes

500

 

Figure 9: The value of


kb/s.

400
v (kb/s)

300

1920

Figure 10: Histogram of residual delays.

200
100
0
-100
-200
200 400 600 800 100012001400
Packet Size (B)

Figure 13: The residual plane of Example 3.

2.3 Application: statistical inference about


long TBFs
Figures 12 and 13 show two examples of the residual
-plane, i.e. the plot
,
,

One block can carry bits of two different IP packets


when the SNDCP segment consists of more than one
packet. In this case none of these packets can be forwarded before all the blocks that carry the segment arrive to the BSC and the corresponding SNDCP layer segment arrives to the SGSN, and when they are forwarded,

@  (  


    

A
4

 


'

A
4

Time

A1 A2 A3 B1 B2 C1 A2 C2
t1
t2
Temporary Block Flow (TBF)

Successful
Retransmission

Erroneous
Block

A
C

9
C

Block Retransmission

one block. In both cases, the value of the residual delay


is the time of one block but negative!
Hence, by looking at the RLC retransmissions that occur in the same TBF we see that the value of the interarrival delay can be as small as the time of 1 block
which, due to the Abis interface, is always 20 ms, and
can be negative. A negative
hence the residual delay
value may thus indicate that there has been at least one
erroneously received block which carried bits of some
earlier IP packet and thus required at least one retransmission. Positive values may in general be due to signalling blocks, block retransmissions, multiplexing of
other TBFs into the same PDCHs, or other more or less
tractable reasons.
Since the LLC layer was not in acknowledged mode
the order of packets may change. Unfortunately, the IP
identier eld was not saved during the measurement of
the data. With that information we could have checked
the order in which packets were originally transmitted
from the MS.
In our theoretical example we can infer that the length
of the TBF that carried packets ,
and
has been
at least the duration of 8 blocks (8 20 ms = 160 ms)
instead of the optimal duration of 7 blocks (140 ms). In
real data we can also make some statistical conclusions
of how the bit rate that, for example, a TCP connection
sees is uctuating. Note that we can recognize only those
TBFs that have successfully carried packets.
Moreover, we can reconstruct the TBF ow of our theoretical example up to some level: we know the positions
,
and
and we know that
and
of blocks
were before
,
was before
and
was before
. If we could know the original order in which the
packets were transmitted from the MS, we could reconstruct our theoretical TBF to even higher level. The observed long TBFs could also be reconstructed up to a
surprisingly high level. If the QoS proles are taken in
use this might make it easier for the operator to detect
the possible QoS level that the user gets for the upstream
trafc.

the measured packet interarrival time can be almost zero.


Assume, for example, that two packets are multiplexed
into one segment. In this case the residual delay of the
rst packet is the same as its size dependent delay, but
negative:
. The size dependent delay of the
latter packet is the sum of the size dependent delays of
both packets. Figures 6 and 7 contain also examples of
this phenomenon. The denition of the residual delay is
thus actually more complicated than in the previous section. In some other sessions there are interarrival times as
small as 1
since the nominal precision of time stamps
was six decimals. This seems to be quite a common phenomenon. In the upstream case it is a property of the
corresponding MS.
Due to block retransmissions, and when the LLC layer
is not in acknowledged mode, the rationale of multiplexing many packets into one segment is more than
questionable since it may delay several packets simultaneously and signicantly. The performance of GPRS
would probably be improved if blocks of such a multiplexed segment would be sent always using CS-1, i.e.
maximizing reliability, like in the case of signalling
blocks.
Anyhow, several IP packets within one LLC frame do
not explain all of the negative values of the residual delays. Consider the case where three packets, , and ,
are sent from an MS in this order and within one TBF in
one PDCH. For simplicity we assume rst that one block
carries bits of one IP packet only. The packet is carried by three blocks whereas packets and are each
carried by two blocks. Assume that the second block
carrying bits of user packet , denoted by
in gure
14, is not received correctly at the BSS and a retransmission is done as soon as possible when the MS receives a
negative acknowledgement. At the BSC the LLC frames,

t3

Figure 14: A theoretical example of a single block retransmission that occurs within the same TBF that uses
one PDCH.
which in this case can be thought of as the IP packets, are
reassembled and relayed to the SGSN and the GGSN in
the order B, A and C. The data that would be measured
in this case,
,
and
, give two pairs
and
. The interarrival
delay
is the time of two blocks, and
is the time of

Optimal measurements

The data we used in this paper was not optimal in the


sense that it was originally measured for a different purpose, for trafc characterization. The quality of time
stamps was good enough to show examples of how the
data analysis procedure should be done. Unfortunately,
this quality was not good enough to go further on. The
nonavailable IP identier eld is one example of information which would have been extremely useful in our
approach. After having done some preliminary studies
with the available data it is easy to plan optimal measurements in the case of GSM/GPRS, when only RLC/MAC
layer works in acknowledged mode. It is clear that if
the LLC layer works in acknowledged mode, the performance of the RLC layer cannot be visible any more.

 

   
   

     



Instead of choosing randomly large sessions or ows at


the monitoring system described in the previous subsection, one could capture trafc from a xed MS with selfgenerated trafc. In this way one could easily arrange
variability in packet sizes by either going through all
packet sizes systematically or selecting the packet size
randomly at the MS. Then one could also control a little the time gap between two successively transmitted
packets. Such a measurement should give enough information for calibrating and testing the passive measurement or monitoring system scenarios. It could also give
some evidence whether the assumption that the bad performance periods in uplink and downlink are often correlated is true or not.
This type of an active measurement is also one of the
rst steps in extending the ideas of the present paper to
the study of WCDMA and WLANs.

3.2 Active measurement

In order to see what kind of service the user IP packets


get we must measure at the point where the IP packet
is completely reassembled. The optimal measurement
for the GPRS analysis shown here should be made from
the SGSN or from the Gn interface between SGSN and
GGSN, see gure 1 at Section 2.1. At the Gn interface the trafc is tunneled by GPRS Tunneling Protocol
(GTP), 3GPP TS 09.60. Some freely accessible protocol analyzing programs like Ethereal, www.ethereal.org,
can dump the GTP. The trace should have accurate time
stamps and, at least, the user IP packet size, GTP packet
size and IP Identication eld from the user IP packet
header. The transfer rate and the transfer media of the
monitoring interface, the accuracy of time stamps and
the drift of the time stamping clock should be known or
veried beforehand.
The GTP header contains a tunnel identier (TID)
which points out the MM and IP contexts. The interesting result of [3] about user mobility during the data
transfer was based on the MM information which is, and
which has to be, quite detailed when the MS is in ready
state. Hence this type of a passive measurement together
with some simple and fast TBF analysis procedure could
be extended to a monitoring system described in the introduction. The main advantage of this approach is that
from only a few monitoring positions it should be possible to see some upstream performance characteristics of
all active BSSs in the whole network.

knowledged mode. In this specic case we have shown


that it is possible to some extent to detect the behavior of
the Gb, Abis and TDMA based Um interface of GPRS
from the MS originated IP level packet data with accurate time stamps. More precisely, it is possible to detect
the number of PDCHs and channel coding scheme used,
extract long TBFs and detect retransmissions of blocks
within a long TBF. We do not know yet the best way
to make informative statistics about TBFs, which is thus
one item for further study. We also described a method
which could be used to see whether the upstream Gb interface was congested or not. Although we do not yet
have any examples, it should also be possible to detect
whether multiplexing of users at the Um interface has
been the case. Of course, these are statistical results and
valid only with some probability but we believe that this
probability is rather high.
Moreover, increasing knowledge of the corresponding packet radio interface and further development of the
statistical tools will certainly increase the probability of
correct identication of the above mentioned phenomena. They will also allow to extend the scope of spectroscopy to other packet radio interfaces. A natural concept for further work is to study whether a similar approach, namely the analysis of the empirical distribution
, yields interesting results on the WCDMA based
UMTS and the CSMA based WLANs.
Acknowledgement. The author is grateful to Andre
Broido from CAIDA for his advice and support all the
way from the very beginning, when the rst ideas had
just came up to the authors mind, up to careful reading
and commenting of the draft. Thanks to Research Professor Pertti Raatikainen from VTT for patiently answering all my very elementary questions about multiplexing hierarchies and teaching me how to make intelligent
guesstimates about them. Thanks to Research Professor
Ilkka Norros from VTT for a careful reading of the nal
draft.

 
D 

3.1 Passive measurement

References
[1] A. Broido, R. King, E. Nemeth, and kc. claffy,
Radon spectroscopy of packet delay, in Providing
Quality of Service in Heterogeneous Environments,
J. Charzinski, R. Lehnert, and P. Tran-Gia, Eds. ITC18, September 2003, vol. 5a of Teletrafc Science
and Engineering, pp. 419428, Elsevier.

4 Conclusions and further work

[2] Kilpi J., A portrait of a GPRS/GSM session, in


Providing Quality of Service in Heterogeneous Environments, J. Charzinski, R. Lehnert, and P. Tran-Gia,
Eds. ITC-18, September 2003, vol. 5a of Teletrafc
Science and Engineering, pp. 389398, Elsevier.

In this paper we have analyzed sessionwise IP packet interarrival time distributions taking into account packet
sizes in the case where the IP packet trace was measured
after the GSM/GPRS packet radio interface. In our analysis it was crucial that only the RLC layer worked in ac-

[3] R. Kalden, T. Varga, B. Wouters, and B. Sanders,


Wireless Service Usage and Trafc Characteristics
in GPRS networks, in Providing Quality of Service in Heterogeneous Environments, J. Charzinski,
R. Lehnert, and P. Tran-Gia, Eds. ITC-18, September

2003, vol. 5b of Teletrafc Science and Engineering,


pp. 981990, Elsevier.
[4] R. Chakravorty and I. Pratt, Performance Issues
with General Packet Radio Service, Journal of
Communications and Networks (JCN), vol. 4, no. 2,
pp. 266281, December 2002.
[5] 3rd Generation Parnership Project (3GPP), European Telecommunications Standards Institute
(ETSI), ETSI, http://www.etsi.org/.
[6] S. McCanne and V. Jacobson, The BSD Packet
Filter: A New Architecture for User-level Packet
Capture, in 1993 Winter USENIX Conference, San
Diego, CA., January 25.-29. 1993.
[7] J. Cleary, S. Donnelly, I. Graham, A. McGregor, and
M. Pearson, Design Principles for Accurate Passive Measurement, in Passive and Active Measurement Workshop PAM-2000, Hamilton, New Zeeland,
April 3.-4. 2000.

You might also like