Beruflich Dokumente
Kultur Dokumente
window and segment sizes, the TCP throughput decreases Packet lost rate = 0.0001
0.001
significantly as the packet loss rate increases beyond 10-4. It 0.01
Max. congestion window = 123 segments
can also be observed that at low packet loss rates, the TCP MSS (Max. Segment Size) = 536 bytes
PFTK TCP Reno model
throughput decreases rapidly as the values of round-trip delay 7
10
RTT increases. As the packet loss rate increases, it becomes
the limiting factor for the TCP throughput, even when RTT is
than 10-4 while keeping the round trip delay as low as possible
for the best performance in terms of TCP throughput.
Effects of Round Trip Delay and Packet Loss Rate on TCP Throughput 4
10
8
10 0 100 200 300 400 500 600 700 800
Round Trip Delay (ms)
250
of traffic based on IPTD and IPDV objectives are defined. The are: 10-3 to 10-7 for real time applications, 10-5 to 10-8 for non-
upper bounds specified for IPTD range include: 100 ms, 400 real time applications. Further details are given in [9] which
ms and 1 s. Upper bound on IPDV is 50 ms for classes 0 and specified a range of BER and SDU error ratio requirements for
1, but not specified for the other 4 classes. Except for the each of the four ITU traffic classes, i.e., Conversational,
unspecified class 5, upper bounds on IPLR and IPER are 10-4 Streaming, Interactive and Background classes. Assuming bit
and 10-3 respectively. errors are independent, for BER = 10-8, octet error rate is
approximately 8x10-8. Again, assuming independent octet
2) Requirements on error rates in IEEE Std. 802 [6]
errors, the MSDU packet error rate is approximately 1.2x10-4.
The IEEE standard for local and metropolitan area networks: This result is 10 times higher than that for IEEE 802.16.3 in
Overview and Architecture, IEEE Std. 802-2001, defines the the Section 3), while consistent with that in Section 2).
compliance with the family of IEEE 802 Standards. It
describes the relationship of the IEEE 802 standards to the 5) Error rate performance supported by GSM
OSI Basic Reference Model and explains the relationship of The error rate performance of GSM as reported in [1] can
these standards to the higher layer protocols. support a bit error ratio of 10-3, which is then reduced to 10-8
in the nontransparent mode radio link protocol (RLP), at the
Subsection 7.3 of IEEE Std. 802-2001 states that the require expense of variable, additional delay due to retransmissions,
error performance of IEEE 802 LANs and MANs shall be less reducing the user throughput.
than 8x10-8 per octet of MAC service Data unit (MSDU)
B. Video over IP error rate requirements
length. While this error performance has to be accomplished at
the physical layer for wired or optical fiber physical media, it For real-time video signal, the contribution to ITU-T SG13
is allowable for this error performance to be accomplished at [10] has stated that IPLR must be at least 10-5. This is derived
the MAC service boundary in the case of wireless media. based on a BER of 10-9 for typical fiber optic network, and the
worst-case assumptions that the packet size is 1500 bytes and
For example, for an MSDU packet with 1500 octets, the that a bit error caused the whole packet to be lost.
required packet error rate will be approximately 1.2x10-4:
In most cases, User Datagram Protocol (UDP) is used for the
PER ≈ 1500 × 8 × 10 −8 = 1.2 x10 −4
transport of video streaming applications. If the UDP
Note that this value agrees closely with the IPLR requirement
checksum is enabled, a packet may be discarded because of a
specified in ITU-T Y.1541 as discussed in Section 1.
single bit error. Since UDP does not allow a re-transmission of
3) Error Rate Requirements in IEEE 802.16.3 [7] the lost packet, the effects of losing a complete video packet
Error rate requirements for the wireless metropolitan area could result in serious disruption to the video signal. However,
network, IEEE 802.16.3, are specified in the section for the UDP checksum is normally not enabled. In that case, the
quality of service support for different classes of services. The packet with bit errors will be received together with the error
maximum BER allowable are 10-6 and 10-4 respectively for bits. Depending on the location of the error bits, the effects of
full and standard quality telephony respectively. The the lost may be tolerable to the user at the receiving end.
corresponding maximum one-way access delay, as measured C. Requirements for various applications and service classes
at the MAC interface with the upper layers, are 20ms and 40
Some examples of error rate and delay requirements are given
ms respectively. For time critical packet services, the
in the 3GPP technical specifications for services and QoS
maximum BER required by 802.16.3 is 10-6, with maximum
architecture [8], [9] & ITU-T recommendations G.114 [13],
one-way access delay of 20ms, whereas the BER requirement
G.1010 end-user multimedia QoS categories [14]. These are
for non-time critical services is 10-9, without requirement on
summarized in the following subsections.
the maximum delay. The corresponding octet and packet error
rates can be computed as follows. Assuming bit errors are 1) Real-time Conversational classes
independent, for BER = 10-9, For conversational voice, acceptable performance for
Octet Error Rate ≈ 1 − (1 − BER ) = 8 × 10 −9
8
maximum frame erasure rates is 3%. The requirement is also
Thus, for a MSDU with 1500 octets, the packet error rate, dependent on the error resilience of speech codec. The
assuming independent octet errors, is approximately 1.2x10-5. requirements for AMR codec is shown in Table 1.
Comparing the error performance requirement for 802.16.3 For videophones, the delay requirement should be similar to
non-time critical packet services with that specified in IEEE that of conversational voice, with additional requirement to
Std. 802-2001 in subsection 2), this exceeds the latter the limits of lip-synch. The maximum acceptable FER is 1%.
requirement by approximately 10 times.
For interactive games, one-way delay value of 250 ms has
4) Error rate performance supported by 3GPP been proposed in [8]. Detail studies on multiplayer network
The ranges of error rate performance that are supported by gaming have been reported in [11]. The studies showed that
3GPP have been specified as part of the QoS requirements in the range of acceptable maximum round-trip time varies
[8]. The ranges of BER that the network is required to support between 200 ms and 40 seconds depending on the type of
251
games, the experience level of the players. Round-trip time is IV. LATENCY OR PACKET TRANSFER DELAY
the total time for a packet to travel from the server to the client
A. Statistics of Internet delay
and the time used to send a reply back from the client to the
server. Delay tolerance is based on the nature of the Internet Some studies have been performed regarding the statistics of
game: Action (Conversational), Real-time (Interactive) or Internet delay for local and International routes [12]. The
Turn-based (Interactive/Background), as shown in Table 1. probability distribution function of one-way Internet delay is
The residual BER requirements correspond to 10-2/10-3, found to follow the shifted Gamma distribution. The mean
10-4/10-6 SDU error ratio depending on the SDU size. delay values for local and International routes are about 10 ms
and 110 ms respectively.
For Two-way control telemetry, an one-way delay limit of 250
ms is proposed in [8], because of the importance of receiving Delays in the public IP network may not be controllable nor
this category of data in a timely manner. In addition, there guaranteed until QoS capabilities such as DiffServ or IPv6 are
should be “0” information loss at the receiving end. widely deployed. However, as the importance of including an
efficient IP network increases, service providers may route the
Table 1 Error rate and delay requirement for various application types
data traffic through private IP network, which may provide
Application Type/ Bit Error Rate End-to-end Delay Source ample capacity to avoid the packet delays due to congestion.
Bit rate As an example, some statistics on packet round-trip time
-4
Conversational Class 1: 10 Preferred: 0-150 ms G.114
voice (AMR codec)/ Class 2: 10-3 Acceptable: G.1010
(RTT) have been collected through Sprintlink Looking Glass
4.75-12.2 kpbs Class 3: 10-2 150-400 ms tools. The average RTT results for a ping packet with a size of
Multiplayer Gaming Action: 10-3 <= 80 ms [11] 100 bytes are summarized in Table 2. In general, the delay in
Interactive: 10-5 250 ms [8] sending packets within the same state is about 15 ms; between
/6x10-8 US East and West coasts is about 75 ms; between the US East
Turn based: 6x10-8 < 40 s
Two-way control ~0 (end-to-end) < 250 ms [8]
coast and Asia/Australia is about 210ms/225 ms, and between
telemetry the US West coast and Europe is about 160 ms.
Web browsing, E- ~0 (end-to-end) Interactive: 2-4 s /
Table 2 Average round-trip time from Sprintlink looking glass
Commerce, Email 0.5s (recommended)
-2
Audio streaming 10 (packet loss) 10 s [14] Stockton,CA Hong Kong Sydney Hamburg
Video Streaming 10-6 – 10-3 150 – 400 ms New York 75 ms 213 ms 225 ms 86 ms
(MPEG4)/ Anaheim, CA 15 ms 158 ms 168 ms 158 ms
24-128kps
Fax, SMS 10-6 30 s [14] B. TCP Round Trip Delay
For a single user link, it can be observed from Figure 1 that it
2) Interactive class would be close to ideal to have RTT of 10 ms in order to
Applications that fall into this class include voice messaging achieve TCP throughput of about 40 Mbps at IPLR of 10-4.
which has similar error rate requirement as that of the While a high user throughput is desirable for applications that
conversational class, but the delay can be up to a few seconds. require block transfer of significant amount of data, it may not
Other applications in this class include: web browsing, E- be realistic to have RTT of 10ms due to the constraints on the
Commerce and Email, as shown in Table 1. PHY and MAC layers, fairness considerations in the
scheduling algorithm in a multiple access network. Moreover,
3) Streaming class the delay in IP network as discussed in the above section
Audio streaming is mainly an one-way application from the indicated that the RTT exceeds 10 ms in most cases.
server to the user. Specific contents of the audio stream may
include high quality music or broadcasting. Thus the error rate V. IETF QoS CLASSIFICATIONS
requirement is more stringent than conversational class but the The Internet Engineering Task Force (IETF) has defined a
delay requirement is more relaxed in comparison. service classification [15] for DiffServ in order to support the
performance requirements for the end user as recommended
The situation with video streaming is similar to audio by ITU-T [4, 5, 14]. The service classes are shown in Table 3
streaming. For MPEG-4 video, the BER ranges between 10-6 with the associated DiffServ Code point associated with each
and 10-3 with significant degradation for the latter. Similar class. It can be observed that further granularity in
requirements apply to the transmission of still image. The classification is provided in the Assured Forwarding (AF)
error tolerance is mainly dependent on the encoding and class.
compression formats. The requirements for audio and video VI. RECOMMENDATION
streaming are shown in Table 1.
As it is desirable for a mobile broadband wireless access
4) Background class system to support a variety of traffic classes, it is necessary to
There is no stringent requirement on delay for the background define the latency and packet error rates performance
class of services. The requirement for fax, and low priority requirements, in order to meet the end-user QoS requirements
transaction services such as short message services (SMS) can for the various applications, as recommended by ITU G.1010,
be found in Table 1. Y.1541. These traffic classes should be mapped to the
252
appropriate QoS classes as defined at the IP network. control are located within reasonable network delay.
Depending on the network configuration, the air interface (AI) Furthermore, when the network delay is the dominating factor,
should support appropriate latency and packet error rate the end-to-end delay is less sensitive to the delay at the AI .
performance targets associated with each traffic class, such
VII.CONCLUSION
that the end-to-end QoS requirements for these applications
can be achieved. Based on the PFTK model for the TCP flow control protocol,
a recommendation for the error rate and latency requirements
Table 3 Service classes, DiffServ code point mapping defined by IETF
for the air interface of a mobile broadband wireless access
Service Class DSCP name Application Example system that supports various IP based applications have been
Administrative CS7 Heartbeats derived, taking into consideration the corresponding
Network Control CS6 Network Routing
requirements for portions of the end-to-end IP network.
Telephony EF, CS5 IP Telephony
Multimedia AF41-AF43 Video Conferencing, ACKNOWLEGMENT
Conferencing Interactive Gaming
Multimedia AF31-AF33, CS4 Broadcast TV, Pay per View, The author thanks Mr. Jim Landon for providing the reference
Streaming Video Surveillance materials and his valuable comments and discussions.
Low Latency AF21-AF23, CS3 Client/Server transactions,
Data peer-to-peer signaling REFERENCES
High Throughput AF11-AF13, CS2 Store & Forward applications, 1. G. Xylomenos, F. Polyzos, P. Mahonen, M. Saaranen,
Data Non-critical OAM&P
Standard DF (CS0) Undifferentiated applications ‘TCP Performance Issues over Wireless Links’, IEEE
Low Priority CS1 Any flow that has no BW communications Magazine, April 2001.
Data assurance 2. Jean Walrand, ‘Communication Networks-A first course’,
McGraw Hill, 1998
In the case of IETF DiffServ, a recommendation on the 3. J. Padhye, V. Firoiu, D. Towsley, J. Kurose, “Modeling
requirements for the AI is shown in Table 4, based on the TCP Reno Performance: A Simple Model and Its
following assumptions: Empirical Validation”, IEEE/ACM Trans. On
Networking, Vol. 8, No.2, April 2000.
i) MAC SDU size = 1500 bytes
4. “Network performance objectives for IP-based services”,
ii) Network Delay = 20 ms or 100 ms
ITU-T Y.1541, May 2002.
iii) Size of IP packet: 1 MAC SDU / IP packet
5. “Internet Protocol Data Communication Service – IP
iv) For UDP traffic:
packet transfer and availability performance parameters”,
End-to-end One-way Delay ≈ AI Delay + Network Delay
ITU-T Y.1540, Jan 2002.
Table 4 Recommendation on requirements for the air interface delay 6. “IEEE Standard for local and metropolitan area networks:
and target packet error rate Overview and Architecture”, IEEE Std. 802-2001, March
IETF Service class Transport End-to-end Network Max. AI MAC SDU
8, 2002.
Protocol One way Delay Delay Packet 7. “Functional Requirements for the 802.16.3
Delay (s) assumed (ms) (ms) Error Rate Interoperability Standard”, IEEE 802.16.3-00/02r4,
Administrative TCP 0.10* 10 10 10-6 September 22, 2000.
Network Control
8. 3GPP TS 22.105 V 6.2.0 (2003-06), Technical
[8, 9,14] TCP 0.25* 10 20 10-6 Specification Group Services and Systems Aspects,
Telephony [8,14] Service Aspects; Services and Services Capabilities.
(Voice over IP) UDP/ RTP 0.15 100 50 3 x 10-2
Multimedia 9. 3GPP TS 23.107 V 5.10.0(2003-09), Technical
Conferencing [8,14] UDP/ RTP 0.15 100 50 10-3 Specification Group Services and Systems Aspects,
Multimedia
Streaming [8,14] UDP/ RTP 0.25 100 150 10-3
Service Aspects; QoS concept and architecture(Release 5).
Low Latency Data 10. Video Performance requirements for IP performance
[8,14] TCP 2* 10 20 10-5 recommendations,ITU-T SG13 D.228(WP4/13),Jan 14,02.
(E-transactions)
High Throughput 11. Multiplayer Game Performance over Cellular Networks,
Data e.g., Email TCP 6* 10 20 10-4 Ver. 1.0, Forum Nokia, Jan 20, 2004
[4, 8, 9,14]
Standard: e.g., FTP
12. Statistics of One-Way Internet Packet Delays, A. Corlett,
[4, 8, 9,14] TCP 10* 10 30 10 -4 D. Pullin, S. Sargood, 53rd IETF, March 18, 2002.
Low Priority Data 13. ‘One-way transmission time’, ITU-T G.114, May 2003
[4, 8, 9,14] TCP 100* 10 40 10-3
*Time required to transfer the all the packets for the email or file.
14. ‘End-user Multimedia QoS Categories’, ITU-T G.1010,
Nov. 2001
As the network delay varies greatly across different networks, 15. F. Baker et. al., IETF Draft “Configuration Guidelines for
under different loading conditions, the values shown in the DiffServ Service Classes draft-baker-diffserv-basic-
table are selected to emphasize the significance of AI delays classes-02”, Feb 13, 2004.
under the selected network conditions [16]. In addition, the 16. ‘A new option proposed for 802.20 requirements on
recommendation is based on the assumption that the data latency and packet error rates’, IEEE C802.20-04-55,
servers for time critical services in the administrative, network May 10, 2004.
253