Cellular Backhauling Optimization

1
Cellular Backhauling Optimization

• Yaakov (J) Stein
• Chief Scientist
• RAD Data Communications, Ltd.

2
Cellular communications

• To most people cellular communications means only
  the air interface
• This is the Radio Frequency link between MS and
  BTS

3
Cellular networks

• But there is a lot more to the cellular network
  than that !
• Using GSM (2G) terminology
• All the Base Transmitter Stations are to Base
  Station Controllers
• The BSCs are connected to Mobile Switching
  Centers
• MSCs are interconnected,
• and also connected to the Public Switched
  Telephony Network

4
Cellular backhauling

• We (informally) call all of the network except
  the air interface
• the cellular backhaul network
• Backhauling of 2G cellular traffic uses TDM
  (E1/T1) links over
• Copper
• Fiber
• Microwave
• Due to rapid worldwide increase in cellular
  penetration
• backhauling is one of the hottest topics in the
  telecommunications industry
• To reduce operational expenses, cellular
  operators want to
• reduce bandwidth consumption
• migrate to (less expensive) Packet Switched
  Networks (IP/MPLS/Ethernet)
• employ less expensive transport types, for
  example
• Metro Ethernet Networks
• DSL links
• Reduction of bandwidth (optimization) for 2G GSM
  is the main topic of this talk

5
Cellular backhaul optimization

• Voice traffic is already compressed by the mobile
  station
• So why can cellular traffic be optimized at all ?
• TDM transport mechanisms can not reduce bandwidth
• standard user traffic (TRAU) formats are
  extremely inefficient
• nonactive user channels are sent
• silence/idle frames in active channels are sent
• signaling channels (HDLC-based) are inefficient
• data can be compressed by lossless data
  compression
• additional mechanisms (e.g. stronger compression)
  may sometimes be used

6
Cellular backhaul transport

• When TDM transport (e.g. E1 links) is used
• optimization enables use of fewer E1s
• to carry the same amount of user traffic
• reduced operational expense at dense portions of
  network
• however, compressed traffic formats are not
  standardized
• When TDM transport is replaced with Packet
  Switched Networks
• service less expensive to begin with
• service often charged by bandwidth used
• optimization enables using only the minimum BW
  needed
• operational expense reduced

7
GSM 2G architecture

• GSM formally separates the Public Land Mobile
  Network into subsystems
• and defines the interfaces / protocols between
  each two pieces of equipment
• A-type interfaces carry the voice traffic in the
  backhaul portion of the network
• A interface is a standard TDM link divided into
  64 kbps timeslots
• Abis interface connects the BTS to the BSC and
  carries FR or HR channels
• Ater interface connects the BSC to the MSC and
  carries FR or HR channels
• A-type interfaces also carry control information

8
Carrying A-type interfaces over PSN
Abis interface
Abis interface
pseudowire (PW)
TDM
TDM
cellopt GW
cellopt GW
BTS
BSC
PSN

• Cellular operators can transport Abis/Ater over
  PSNs instead of TDM
• To do this without forklift upgrade of their
  equipment to 3G
• they can use pseudowire (PW) technology
• A PW emulates a native service by building a
  tunnel through the PSN
• Bandwidth reduced as compared to TDM
• with optimization, bandwidth can be further
  reduced

9
Voice channels

• Although over time new services were added
• Fax
• Short Message Service
• Multimedia Message Service
• Wireless Application Protocol
• Internet and WWW access
• Video streaming
• the cellular network was originally designed for
  voice traffic
• A GSM transmitter segments voice into 20
  millisecond frames
• And applies compression to place voice traffic
  into one of two channel types
• Full Rate channel - 16 kbps 2 bits every 1/8000
  sec. 320 bits per 20 ms.
• Half Rate channel - 8 kbps 1 bit every
  1/8000 sec. 160 bits per 20 ms.
• There are various compression algorithms
• Full Rate codec - 13 kbps (FR channel)
• Enhanced Full Rate codec - 12.2 kbps (FR channel)
• Half Rate codec - 5.6 kbps (HR channel)
• Adaptive MultiRate - 4.75, 5.15, 5.9, 6.7, 7.4,
  7.95 (HR or FR channels)
• - 10.2, 12.2
  kbps (FR channel)

10
TRAU frames

• The compressions and format conversions in the
  network are performed by the
• Transcoder and Rate Adaptation Unit
• Information on the A bis and A ter interfaces is
  encoded in TRAU frames
• TRAU voice frames represent 20 ms. of audio
• FR channels - TRAU frames are 320 bits 40 bytes
• HR channels - TRAU frames are 160 bits 20 bytes
• The TRAU frames are transported over FR and HR
  channels

idle 01 alarm 00
Note that a full E1 (2 Mbps) must be used even
when there are very few channels
11
TRAU framing
The TRAU frames have a specific frame structure
that must be detected For example, this is the
framing of the generic FR (40 byte) TRAU frame
And this is the generic HR (20 byte) TRAU
frame Note there are other frame formats
as well
x bits are data / control and are not part of
the framing T bits are for time alignment
(justification)
12
Abis signaling channels

• It is very important not to delay or corrupt
  special signaling channels
• Ater signaling channels are based on SS7
• Abis signaling channels are not completely
  standardized
• each equipment vendor has its own signaling
  format
• Abis Signaling channels can be
• 16 kbps (2 bits per TDM frame)
• 32 kbps (4 bits per TDM frame)
• 64 kbps (a full 8 bit TDM timeslot)
• Signaling is usually HDLC based, with a frame
  format
• The frame between flags (7E hex) is bit-stuffed
• Between frames there may be flags or other
  filling

13
Backhauling data

• User data can be transported over the Abis
  interface in various ways
• Low rate data (up to 9.6 or 14.4 kbps) is
  transported in TRAU frames
• Intermediate rates (up to 114 kbps) are available
  via GPRS (2.5G)
• Higher rates (theoretically up to 384 kbps) via
  EDGE (2.75G)
• GPRS / EDGE are carried over G-type interfaces
• which may share the same TDM link as A-type
  interfaces
• GPRS/EDGE bandwidth allocation may be dynamic
• it takes over bits not used by the A-type
  interfaces
• In 3G networks data can be much higher rate (over
  2 Mbps, e.g. 10 Mbps)
• carried over I-type interfaces
• that use separate transport media

14
GSM 2.5G architecture (GPRS/EDGE)

• The first high-speed GSM data (WAP, PTT, MMS,
  WWW) service was
• the Generalized Packet Radio Service
• It provides 56 kbps - 114 kbps packet data for IP
  communications
• The air interface is enhanced, but won't be
  discussed here
• To the GSM backhaul architecture is adds
• Serving GPRS Support Node
• Gateway GPRS Support Node
• G interfaces
• The next stage is Enhanced Data for Global
  Evolution (AKA Enhanced GPRS)

15
3G architectures

• In initial 3G releases Iu interfaces are based on
  ATM (Iu-CSAAL2, Iu-PSAAL5)
• In the final phases, the network becomes IP
• and the protocols become VoIP
• At that point the window of opportunity for
  optimization closes

16
1st challenge - channel detection

• Voice/data/signaling information appears at
  various places in the frame
• Were we to understand the proprietary signaling
• we would know where to look for the various
  channels
• but this signaling is vendor-dependent
• and the formats are not always known
• So we need to employ an intelligent
  detector/classifier/deframer
• detect channel framing and return field positions
• classify channel as voice/data/signaling/idle/unkn
  own
• maintain relative synchronization
• Matching framer at egress needs to recreate the
  original frames

17
Channel detector/classifier

• This detector/classifier needs to continually
  scan all
• 1-bit positions for HR TRAU frames
• even aligned 2-bit fields for FR TRAU frames
• even aligned 2-bit fields for HDLC
• nibble-aligned nibbles for HDLC
• byte-aligned octets for HDLC
• fields of idle bits
• anything else
• and then return the identifications and positions
  found
• Unidentified non-idle information must be
  reliably transported
• The processing involves
• searching for specific bit combinations
• performing bit correlations
• and is extremely computationally intensive
• Can be performed by a DSP with good bit-oriented
  operations

18
2nd challenge - optimization

• Once the various components have been found
• the information needs to be reduced in size and
  reliably transported
• Idle fields need not be sent, often accounting
  for a large BW reduction
• TRAU framing overhead may be removed
• Voice frames marked as silent (DTX) may be
  suppressed
• Voice Activity Detection may be employed to
  suppress silence
• HDLC flags are removed and the contents destuffed
• Data may be compressed
• We will deal with each of these in turn

19
3rd challenge - data compression

• Data is typically transported over cellular
  networks in uncompressed form
• Lossless data compression algorithms, e.g.
• Ziv-Lempel variants
• Huffman codes / arithmetic codes
• Shannon-Fano coding
• Burrows-Wheeler Transform
• Prediction by Partial Match
• can be an effective optimization method
• when there is a significant amount of data
  traffic
• Text data, such as HTML or WML, can be
  significantly compressed
• Compressed video, binary files, encrypted data,
  etc.
• can not be compressed

20
Using data compression

• Many algorithms perform well when there is a lot
  of data
• The problem is that the impact of packet loss
  must be taken into account
• If we compress each packet separately
• there is not enough data for efficient
  compression
• If we keep history from previous packets
• we need to separate flows
• we need to store state
• loss of single compressed packet causes multiple
  packets to be discarded
• DSPs can be exploited to handle data compression
• main limitation - large amount of memory needed
• Need a DSP with efficient bit/byte-oriented
  operations

21
4th challenge - trans-rating

• Audio / video streams are already compressed
• Further compression may not be possible
• However, sometimes there are hard bandwidth
  limits (caps)
• and we must be able to survive short bandwidth
  peaks
• In certain instances trans-rating may be useful
• at the expense of reduced perceived quality
• especially when exceeding the cap is expected to
  be extremely rare
• For example
• change compression rate for AMR family on a
  frame-by-frame basis
• transcode EFR codec down to a lower AMR rates
• and transcode back up at network egress

22
Smart trans-rating

• The simplest (but most computationally intensive)
  way to trans-rate
• is to cascade a decoder and an encoder
• For a particular pair of codecs there may be
  better ways, with
• lower computational complexity
• lower delay
• less perceptual degradation
• For AMR, there are commonalities that may be
  exploited
• However, reserving DSP computational resources
• is usually not economically justifiable
• for a process that will only be used for short
  bandwidth peaks
• Other mechanisms may be more affordable, such as
  smart frame drop

23
5th challenge - smart frame drop

• Sometimes transport traffic bandwidth has a hard
  cap
• If this cap is exceeded, voice frames will be
  discarded
• The TRAU will employ Packet Loss Concealment
  techniques
• that cover up much of the effect
• generally there will still be noticeable impact
  on the user experience
• A smarter technique is smart selective frame drop
  (extended VAD)

24
Smart frame drop

• Instead of dropping randomly chosen voice frames
  
• we can carefully select the frames to drop
• using a criterion of least perceptual quality
  degradation
• The selection can be based on voice parameters in
  the TRAU frame
• without full decoding of the voice coding
• The resulting DSP code
• is codec-dependent
• requires saving of state information per channel
• but does not require large amounts of memory
• The smart frame drop mechanism
• should be tightly coupled controlled to the main
  control function
• so that only the minimal percentage of frames are
  dropped

25
6th challenge - timing recovery

• TDM's physical layer transfers accurate frequency
  (sync) information
• GSM BTSs use the accurate frequency recovered
  from the TDM link to
• generate accurate radio frequencies
• generate symbol timing
• send time offset information to mobile stations
• ensure short handover when moving from cell to
  cell
• CDMA and 3G cellular systems also need accurate
  Time Of Day
• Requirements are stringent
• absolute frequency accuracy must be better than
  50 ppb
• jitter and wander need to conform to ITU TDM
  standards
• 3G stations need time accuracy of better than ?3
  ?s
• 3G TDD mode requires time accuracy of better than
  ?1.25 ?s from UTC
• When replacing TDM links with PWs over PSNs we
  lose timing information

26
Frequency measures

• Frequency needs to be stable and accurate
• and there may be both frequency jitter and wander
• jitter is easy to filter out - the real problem
  is wander

Jitter short term timing variation (i.e. fast
jumps - frequency gt 10 Hz)
Wander long term timing variation (i.e. slow
moving - frequency lt 10 Hz)
27
PSN - Delay and PDV
TDM frequency distribution is based on constant
bit rate Packets in PSNs may be sent at a
constant rate but PSNs introduce Packet Delay
Variation PDV makes frequency recovery difficult
28
Jitter Buffer
Jitter Buffer

• Data from arriving packets are written into a
  jitter buffer
• Once buffer is 1/2 filled, we read from buffer
  and output to Abis interface
• Data is read from jitter buffer at a constant
  rate - so no jitter
• But how do we know the correct rate ?
• How do we guard against buffer overflow/underflow
  ?
• We need a frequency recovery algorithm

29
Frequency recovery

• Packets are injected into network ingress at
  times Tn
• The source packet rate R is constant
• Tn n / R
• The PSN delay Dn can be considered to be the sum
  of
• typical delay d and random delay
  variation Vn
• The packets are received at network egress at
  times tn
• tn Tn Dn Tn d Vn
• By proper averaging/filtering
• lttn gt Tn d n / R d
• and the original packet rate R has been recovered
• Unfortunately, simple averaging would be much too
  slow
• By the time the accuracy would be sufficient, the
  rate would have wandered
• In such cases control loops (PLL, FLL) are
  commonly used
• but the noise is much higher here than in usual
  cases where PLLs are used
• and changing frequency to compensate for
  inaccuracy causes wander

30
Frequency recovery algorithms

• Early solutions relied on
• linear regression
• augmented PLLs
• FLL - PLL hybrids
• More sophisticated implementations exploit
• parameter estimation and tracking
• oscillator modeling
• network modeling
• system separation
• Although the algorithms may be complex
• they run at a relatively low rate (tens of times
  per second)
• and can thus be run on a DSP

31
Summary

• Cellular backhaul optimization enables
• more efficient use of overloaded transport
  infrastructures
• lowering of OPEX
• Cellular optimization is applicable to 2G and
  2.5G networks
• There are many challenges to building an
  operational system
• channel detection, classification, and deframing
• packet-loss-tolerant data compression
• smart trans-rating
• smart selective frame drop
• timing recovery
• DSPs provide a good platform for meeting these
  challenges
• For more information, visit www.RAD.com