A Research Framework for the Clean-Slate Design of Next-Generation Optical Access Dr Kyeong Soo (Joseph) Kim College of Engineering Swansea University 22 August 2011
Jun 20, 2015
A Research Framework for the Clean-Slate Design of
Next-Generation Optical Access
Dr Kyeong Soo (Joseph) Kim
College of Engineering
Swansea University
22 August 2011
Outline
I. Introduction
II. New Comparative Analysis Framework
III. Virtual Test Bed for Experiments
IV. Preliminary Results: Elasticity of Hybrid PON
V. Summary
2
Overview of Proposed Research Programme
Virtual Test Bed for Next-Generation Optical Access (NGOA)
Comparative Analysis Framework
Generating realistic traffic through a complete protocol stack
based on user behaviour models at application & session levels.
Quality of Experience
(QoE)
• A new comparative analysis framework for the user-
perceived performances (as measures of QoE) of candidate
NGOA systems.
• It is based on equivalent circuit rate (ECR) framework &
multivariate non-inferiority testing procedure and able to
take into account the statistical variability in experimental
data and a tolerance for the measure.
• The virtual test bed is implemented as simulation models for both systems under
test (SUT) &supporting environments (i.e., application server, intermediate
routers, user nodes) and run in a large-scale with the help of cloud computing (i.e.,
Amazon elastic compute cloud (Amazon EC2)*).
• It provides a common reference framework for experiments for researchers in
both Academia and Industry in order to properly benchmark candidate NGOA
systems and exchange their results.
Energy-Efficient and Elastic Components and Architectures
New Research Framework for Clean-Slate Design of NGOA
• Based on the proposed research framework,
we will investigate from scratch the
following major issues in components &
architectures:
• System capacity
• Elasticity
• Cost & energy efficiency
• We expect that the results from our study
will be quite different from conventional
ones and, therefore, have many implications
on traffic engineering as well as
architectural designs for the following
reasons:
• Realistic traffic models
• Measures for user-perceived
performances
User Behaviour-Based Traffic Modelling/Generation
* Supported by Amazon Web Services (AWS) in Education Research Grant.
I. Introduction
4
Changing Landscape for NGOA in 10-Plus-Year Time Frame
• Architectural requirements – Higher line rate (10+ Gbit/s)
and per-user capacity
– Lower power consumption
– Elasticity
– Access/metro integration • Higher split ratios (1:128~1024)
• Longer reaches (60~100 km)
– Support of hybrid optical/wireless
• Traffic characteristics – Higher bandwidth
– Higher burstiness
– Peer-to-peer
• Business aspects – Revenue model
– New services/applications • User demands
• Usage behaviour
5
Characteristics of Future Traffic: Higher bandwidth and burstiness!
Slow start
Congestion
avoidance
HTTP (TCP) Traffic
I B B P B B I B … frame
size MPEG-4 Part 2
(I/P/B=6721B/2234B/1186B)
H.264/AVC (I/P/B=5658B/1634B/348B)
Video Traffic*
Cannot fill the pipe
for most of the time
* G. van der Auwera et al., “Traffic characteristics of H.264/AVC variable bit rate video,”
IEEE Comm. Mag., vol. 46, no. 11, pp. 164-174, Nov. 2009. 6
Lessons from Cloud Computing* - 1
– Elasticity
• Ability to add or remove resources at a fine grain and with a small lead time
– Transference of risks of
• Overprovisioning (underutilization)
• Underprovisioning (saturation)
Max. (=peak)
Min.
Avg.
Time
Demand
* M. Armbrust et al., “Above the clouds: A Berkeley view of cloud computing,”
Dept. of EECS, UC Berkeley, Tech. Rep. UCB/EECS-2009-28, Feb. 2009. 7
Lessons from Cloud Computing - 2
– The illusion of infinite computing resources available on demand • Through the construction of large-scale, commodity-computer
datacenters at low cost locations, and virtualization technique
– The elimination of an up-front commitment by cloud users • Companies can start small and increase gradually
– The ability to pay for use of computing resources on a short-term basis as needed
8
Implications on NGOA Architectures - 1
9
Economy of scale through integration*
Network resource as utility
Maximisation of resource sharing
* M. Armbrust et al., “Above the clouds: A Berkeley view of cloud computing,”
Dept. of EECS, UC Berkeley, Tech. Rep. UCB/EECS-2009-28, Feb. 2009.
Implications on NGOA Architectures - 2
Backbone/Core MAN
Access
Access
Residential
Users
Business
Users
Current edge of the OBS/MPLS network
New edge of the OBS/MPLS network (toward ONUs)
New framing &
switching schemes
at UNIs
…
…
Flexible PON Flexible PON with higher elasticity
See the next slides for examples.
10
Remotely Powered & Reconfigurable Remote Nodes*
11
JSDU PPC-9LW
Photovoltaic Power Converter
Fiberer 2x2 MEMS
Optical Latching Switch
* J. H. Lee et al., “A remotely reconfigurable remote node for next-generation access networks,”
IEEE Photon. Technol. Lett., vol. 20, pp. 915, Jun. 2008.
Scheduler
Downstream
Traffic
Queues
. . .
1:M
Passive Splitter
Upstream
Traffic
Queues
. . .
Burst-Mode Receiver
MAC Downstream
Traffic Queue
Upstream
Traffic Queue
Hybrid TDM/WDM-PON OLT
Hybrid TDM/WDM-PON ONU
1:N
AWG
. .
. .
. .
Circulator
Tunable Transmitter
Tunable Receiver
Tunable Receiver
Circulator
. . .
. . .
Tunable Transmitter
Modulator (e.g., RSOA)
Hybrid
TDM/WDM-PON
with Tuneable
Transceivers
1:2
Passive Splitter
12
Key Observations - 1
• Toward more energy-efficient & elastic NGOA
– Performance has long been a major issue in traditional network design, but energy efficiency becomes as important an issue as performance in design of NGOA.
– The elasticity of network is a key to achieving higher energy efficiency as well as lower operation & maintenance cost.
13
Key Observations - 2
• Lack of a comprehensive research framework
– Many candidate architectures and protocols have been proposed, but there is no comprehensive framework to compare their performances in a fair and objective way.
– A new research framework should be able to take into account the energy efficiency and elasticity of architectures and protocols as well as the performances perceived by users (i.e., reflecting QoE).
14
Key Observations - 3
• Importance of experiments in the study of NGOA systems
– Due to the complexity of protocols and the interaction among multiple traffic flows in the study of network architectures, researchers now heavily depend on experiments with simulation models or test beds implementing proposed architectures and protocols, rather than traditional mathematical analyses under simplifying assumptions.
15
Key Observations - 4
• Need of a powerful and flexible test bed for experiments
– To provide the whole protocol stack up to the application layer, measurement of user-perceived performances, and traffic generation based on user behavior.
– To enable researchers and network operators/service providers to test candidate NGOA systems under a realistic and long-term operating environment. • We are aiming to investigate the impact of (at least) daily usage
patterns of residential & business users on network performances.
16
What Should We Do Now?
• Because we are at an early stage of research for long-term NGOA solutions, we need to establish a comprehensive research framework for comparing candidate network architectures and protocols and implement a powerful and flexible test bed for actual experiments under a realistic operating environment.
• Then, we should carry out the investigation of both new (e.g., energy efficiency, elasticity) and old (e.g., performances, cost) issues for candidate network architectures and protocols based on the proposed research framework and test bed.
17
Clean-Slate Design of NGOA
Virtual Test Bed
NGOA Components & Architectures
18
Comparative Analysis
Framework
Revenue & User Behaviour Modelling
18
Need of Clean-Slate Design
• To take into account the fundamental changes of NOGA in 10-plus-year time frame, we need to revisit from scratch the following issues:
– Research framework • Comparative analysis
• Quantification of bandwidth and power consumption
• Traffic generation and performance measures
– Architectures • WDM-PON, hybrid PON, OFDMA-PON, ODSM-PON*.
– Protocols and algorithms • Switching/routing, scheduling, flow control, framing, link adaptation.
– Revenue models • Based on new consumer demands and usage behaviours.
19 * Opportunistic and dynamic spectrum management PON
Our Approaches
• We will design from scratch and investigate the issues of energy efficiency, elasticity, and cost based on user-perceived performances under a realistic & long-term operating condition.
• Our approaches will give us new insights on critical issues in NGOA architectures and protocols including – Quantification of network capacity based on user-perceived performances.
– Evaluation of energy efficiency , elasticity, and cost of candidate systems based on the (statistically) equivalent user-level performances.
– Investigation of merits/demerits of shared & dynamic architectures with respect to dedicated & static architectures.
20
II. New Comparative Analysis Framework
21
Motivation
• Due to the complexity of protocols and the interactive nature of traffic involved in the study of network architectures, researchers now heavily depend on experiments with simulations or test beds.
• Due to the shift toward experiments, comparison procedures should be able to take into account the statistical variability in measured data from the experiments.
• Measures for the comparison should be user-oriented (i.e., QoE-based) and multiple measures should be compared together in an integrated way.
22
Issues in Comparison - 1
• The comparison between two delay curves shown in the figure seems straightforward at first glance.
• In fact, it is not straightforward when we consider the statistical variability in measured data. – A statistical approach is needed!
23
Load
Delay
System A
System B
x
24
Issues in Comparison - 2
• How can we compare multiple performance measures of two systems in an integrated way? – Can we say that the system A is
at least as good as the system B? • If so, on what basis?
• How can we relate network-level performances to user-perceived performance (i.e. QoE)?
25
Measure System A System B
Delay 10 ms 9.5 ms
Throughput 100 MB/s 97.6 MB/s
Packet Loss Rate
5 % 7 %
Existing Works on Comparison Framework
• Equivalence Testing1
– Frequently used in Medicine and Biology for the establishment of the equivalence (often called bioequivalence) between two different clinical trials or drugs through statistical hypothesis testing.
• Equivalent Circuit Rate (ECR)2
– A measure for shared packet access network, which specifies the rate of a dedicated connection that is equivalent to the shared system in terms of user-perceived performance (i.e., web page delay).
– Currently state-of-the-art in networking research.
26
1. R. L. Berger and J. C. Hsu, “Bioequivalence trails, intersection-union tests, and equivalence confidence sets,”
Statistical Science, vol. 11, no. 4, pp. 283-319, 1996.
2. N. K. Shankaranarayanan et al., “User-perceived performance of web-browsing and interactive data in HFC cable
access networks,” Proc. of ICC’01, vol. 4, Jun. 2001, pp. 1264-1268.
App.
Server
App.
Server
RB
RB
Reference Architecture
R = min(RF, RD) ( 1.0)
ONU 1 User 1
User n
…
RF
RD
RD
Candidate Architecture
ONU 1
ONU N
…
User 1
User n
…
User 1
User n
…
RU
RU
OLT
OLT
RU
Comparative Analysis Framework based on ECR
27
Original ECR Calculation Procedure
28
Simulation with reference model (Point-to-point)
Build a function f(R)=Dw
given the # of sessions
Simulation with candidate model
(TDM-PON, Hybrid PON, …)
Find Dw
given the # of sessions
Solve f(ECR)=Dw
* R: Access line rate
* DW: Web page delay
Issues in Original ECR Framework
• Use of a single performance metric
– HTTP traffic alone cannot provide enough load for the NGOA with 10+ Gbit/s capacity.
– A broad spectrum of applications/services for future NGOA needs to be covered.
• No systematic comparison procedure
– No procedure is provided for the comparison of simulation results.
29
New Comparative Analysis Framework
• Construct a new framework extending the original ECR framework as follows:
– What to compare • Multiple user-perceived performances covering broad spectrum of
applications/services for future NGOA
• Percentile (including average and maximum) of performance measures
– How to compare • Comparison of a single performance measure: Based on non-
inferiority testing (at least as good as; one-sided variant of the equivalence testing).
• Integration of multiple single-performance comparisons: Based on intersection-union testing (IUT).
30
Equivalence or Non-inferiority? - 1
• Equivalence test is based on “acceptable difference” limits as follows:
– 𝐻0: 𝜇𝑅 − 𝜇𝐶 ≤ 𝜃𝐿 or 𝜇𝑅 − 𝜇𝐶 ≥ 𝜃𝑈
– 𝐻𝐴: 𝜃𝐿 ≤ 𝜇𝑅 − 𝜇𝐶 ≤ 𝜃𝑈
31
Equivalence or Non-inferiority? - 2
• However, do we really need the two-sided criterion in comparing network architectures based on performance measures (i.e., QoE)?
– “At least as good as (i.e., non-inferiority)” would be more appropriate criterion here.
– So the hypotheses in this case are:
• 𝐻0: 𝜇𝑅 − 𝜇𝐶 ≤ 𝜃 (or 𝜇𝑅 − 𝜇𝐶 ≥ 𝜃)
• 𝐻𝐴: 𝜇𝑅 − 𝜇𝐶 > 𝜃 (or 𝜇𝑅 − 𝜇𝐶 < 𝜃) – Direction of inequality depends on a measure (e.g., delay,
throughput).
32
Extended ECR Calculation Procedure - 1
• First, obtain measures of the user-perceived performances for applications/services (i.e., 𝑀1, ⋯ ,𝑀𝑁𝑀
) of the reference architecture for the
line rates of 𝑅1, ⋯ , 𝑅𝑁𝑅 where 𝑅1 = 𝑚𝑖𝑛 𝑅𝐹 , 𝑅𝐷 ,
𝑅𝑁𝑅> 0, and 𝑅𝑖 > 𝑅𝑗 for 𝑖 < 𝑗.
– 𝑁𝑀 and 𝑁𝑅 denote the number of performance measures adopted and the number of values for R (i.e., 𝑅𝑖’s) used for comparison, respectively.
33
Extended ECR Calculation Procedure - 2
• Second, using the procedures in the next slides, find a value for the line rate of the reference architecture for which the measures of the candidate architecture are statistically non-inferior to those of the reference architecture.
– The null and the alternative hypotheses of the non-inferiority testing for measure 𝑀𝑖 are given by
𝐻0: 𝜇𝑖,𝑅 − 𝜇𝑖,𝐶 ≤ 𝛿𝑖𝐻1: 𝜇𝑖,𝑅 − 𝜇𝑖,𝐶 > 𝛿𝑖
where 𝜇𝑖,𝑅 and 𝜇𝑖,𝐶 denote population means of 𝑀𝑖 for the reference and the candidate architectures, respectively, and 𝛿𝑖 represents the tolerance for the measure 𝑀𝑖.
34
ECR Calculation Procedure Based on Multivariate Non-inferiority Testing
Start
i = 0
Multivariate non-inferiority testing
(of candidate w.r.t. reference with the line
rate of Ri)
result = pass?(Non-inferior?)
i < NR?
No
i = i + 1
Yes
Failed
No
ECR = Ri
Yes
End
• NR: Number of values for R (i.e., Ri’s) used for comparison
35
See the next slide.
Multivariate Non-inferiority Testing Based on Intersection-Union Test (IUT)
Start
i = 0
Non-inferiority testing for measure i
(of candidate w.r.t. reference)
Reject H0?(Non-inferior?)
i < NM?
Yes
i = i + 1
Yes
result = pass
No
result = fail
No
End
• NM: Number of performance measures adopted
36
Benefits of a New Comparison Framework
• We can present the combined performance (i.e., ECR) of a system in a more compact way, given the configuration.
• We can compare non-traditional measures of systems (e.g., energy efficiency, cost, and the amount of resources) in a fairer way under the configurations (of systems) providing the same ECR.
37
Challenges of a New Comparison Framework
• Measurement of user-perceived performances under a realistic operating environment requires
– Traffic models based on user behaviours for applications/services.
– A test bed to carry out experiments under a realistic operating environment and obtain higher-level measures for user-perceived performances.
• We need massive computational power to compute a test statistic of measures for the ECR calculation.
38
III. Virtual Test Bed for Experiments
39
Why Not Real Test Bed?
• A real test bed is good for the R&D of systems already standardised, but not cost-efficient and flexible enough for the clean-slate design of NGOA with the following features: – Multiple candidate systems are tested and compared.
– A whole protocol stack (up to application layer), together with user behavior models, is used for the investigation of interaction among traffic flows as well as realistic traffic generation.
• We want a common reference framework for experiments to be available for researchers in both Academia and Industry (not physically confined to a certain lab/institution) in order to properly benchmark candidate systems and exchange their results. – Like ns in early days of the study of TCP protocols1,2.
40
1. K. Fall and S. Floyd, “Simulation-based comparisons of Tahoe, Reno and SACK TCP,”
SIGCOMM Comput. Commun. Rev., vol. 26, pp. 5–21, Jul. 1996.
2. L. Breslau et al., “Advanced in network simulation,” IEEE Computer, vol. 33, no. 5, pp. 59-67, May 2000.
Our Solution
• A virtual test bed which will be implemented as simulation models for both systems under test (SUT) and supporting environments (e.g., application server, user nodes) and run in a large-scale with the help of cloud computing.
• Features:
– Simulation models based on OMNeT++/INET framework1.
– Message passing interface (MPI)2 for parallel extension of large-scale simulation.
– Hybrid cloud based on local private cloud and remote public cloud for scalable & cost-efficient computing.
41 1. http://www.omnetpp.org
2. http://www.mpi-forum.org
App.
Server
RB RF
RD
RD
ONU 1
ONU N
…
RTT
User 1
User n
…
User 1
User n
…
RU
OLT ODN
RU
Router Router RB
SNI UNI System Under Test
ODN: Optical Distribution Network
SNI: Service Node Interface
UNI: User Node Interface
RTT: Round-Trip Time
Overview of Virtual NGOA Test Bed
42
Traffic Generation in Existing Work
• Traffic is usually generated at the data link/network layer based on a frame/packet-level traffic models. – Due to the computational
complexity involved with implementation of higher layers and protocols.
– This approach, however, cannot capture the interactive nature of real traffic in actual networks.
43
Application/ Session
Transport
Network
Data Link
Physical
or
Traffic Generation in Our Work
• Traffic will be controlled and initiated by the user (via its behaviour model) and generated through the whole protocol layers (including application/session and transport). – In this way, we can capture the
interactive nature of real traffic in actual networks in the virtual test bed.
– Also, we can take into account traffic patterns for different types of users. • e.g., business vs. residential users
44
Application/ Session
Transport
Network
Data Link
Physical
User
Overview of User (Host) Node
TCP
UDP
Network and Lower
Layers
…
…
…
UNI
HTTP 1
FTP 1
HTTP nh
FTP nf
Video nv
Video 1
User Behaviour
Model
45
User Behaviour-Based Traffic Modelling and Generation
User Behaviour Model
Session-Level Traffic Model
Packet/Frame-Level Traffic Generation
X1 X2 Xn
Y1 Y2 Ym
46
Measurement of User-Perceived Performances
• Capturing user-perceived performances as objective measures of QoE
• Use of session- or video-frame-level metric – Session delay; web page delay for HTTP and file downloading delay for
FTP.
– Decodable frame rate (DFR) for video stream.
• Measurement of percentiles – Mean and maximum as special cases.
47
Session-Level HTTP Traffic Model • A behavioural model for user(s) web browsing* with the
following simplification: – No caching and pipelining
– Adapted for traffic generation at the client side
48 * J. J. Lee and M. Gupta, “A new traffic model for current user web browsing behavior,”
Research@Intel, 2007.
Server
Client
Request for
HTTP object
Request
for embedded
object 1
Response
Parsing Time Reading Time
…
Request
for embedded
object 2
Response to the last
embedded object Request
for HTTP
object
Web page delay (= session delay)
Session-Level FTP Traffic Model • A simple model for user(s) file downloading*:
– Could be considered as HTTP sessions just downloading files.
– The model is for a data transfer connection only. • A connection for control information is ignored.
– Adapted for traffic generation at the client side
49
Server
Client
Request for
a file to download
Reading Time
Response to the
file Request for
a file to download
File download delay (= session delay)
* cdma2000 Evaluation Methodology, 3GPP2 C.R1002-B, 3GPP2 Std., Rev. B, Dec. 2009 .
Streaming Video Traffic Model
• Interface with OMNeT++/INET framework
– Through “UDPVideoStream{Svr,Cli}WithTrace” modules:
• UDP server can handle multiple client requests simultaneously
• Random starting phase for each request
• Wrap around to generate infinite streams
• UDP client records the following performance metrics: – Packet end-to-end delay (vector)
– Packet loss rate
– Frame loss rate
– Decodable frame rate (perceived quality metric)
50
Implementation of Virtual Test Bed
GPON Test Bed
Private Cloud
Cloud Computing 51
Hybrid Cloud for Seamless Development and Running of Programs
• We will build a hybrid cloud integrating a small-scale, private cloud and a large-scale public cloud (Amazon EC21)
– An identical environment for both the development of programs at a local private cloud and the run of them at a remote public cloud without change of code
– Based on open-source solutions
• OpenNebula (distributed VM management)
• StarCluster (clustering and load balancing)
1. Currently supported by Amazon Web Services (AWS) in Education Research Grant.
2. R. S. Montero, “Scaling out computing cluster with Amazon EC2: Hands on!,”
ESAC Grid Workshop, Madrid, Spain, Dec. 2008.
2
52
IV. Preliminary Results: Elasticity of Hybrid PON
53
System Model – Dedicated Reference • Number of ONUs (N): 1
• Number of hosts per ONU (n): 1, 2, …
• Access rate (RD = RF): 1, 10 Gbit/s
• UNI rate (RU): 10 Gbit/s
• Backbone rate (RB): 1 Tbit/s (future standard or MUX of 100-Gbit/s links)
• Round-trip time (RTT): 10 ms (including 600 µs RTT in 60-km PON)
54
App. Server
ONU 1 RD = RF
Host 1
Host n
… Backbone
RTT
RB
RU
System Model – Hybrid PON • Number of ONUs (N): 16
• Number of Hosts per ONU (n): 1, 2, …
• Distribution/Feeder Rates (RD, RF): 10 Gbit/s
• UNI Rate (RU): 10 Gbit/s
• Number of Transceivers (NTX, NRX): 1, 2, …
• Backbone Rate (RB): 1 Tbit/s
• Round-Trip Time (RTT): 10 ms
55
RF App. Server
ONU 1
ONU N
…
RD
RD
Host 1
Host n
…
Host 1
Host n
…
RTT
RB
OLT
RU
TX, RX
Number of Sessions per Each Traffic
• nh = nv = 1
– Assume that a user can watch and interact with at most one video channel and one web session simultaneously at any given time. • As far as user perceived (interactive) performance is concerned.
• nf should be kept large to load the high-speed access link
– FTP is usually background process. • This could be HTTP sessions just downloading files!
– Suggest 10 as a starting point. • Cannot find FTP model for 10 Gbit/s link at this time!
56
HTTP Traffic Model -1 Parameters / Measurements Best Fit (Parameters)
HTML Object Size [Byte] / Mean=11872, SD=38036, Max=2 M
Truncated lognormal (=7.90272, =1.7643, max=2 M) Mean=12538.25, SD=45232.98*
Embedded Object Size [Byte] / Mean =12460, SD=116050, Max=6 M
Truncated lognormal (=7.51384, =2.17454, max=6 M) Mean=18364.43, SD=105251.3
Number of Embedded Objects / Mean=5.07, Max=300
Gamma (=0.141385, =40.3257) Mean=5.70, SD=15.16
Parsing Time [sec] / Mean=3.12, SD=14.21, Max=300
Truncated lognormal (=-1.24892, =2.08427, max=300) Mean=2.252969, SD=9.68527
Reading Time [sec] / Mean=39.70, SD=324.92, Max=10000
Lognormal (=-0.495204, =2.7731) Mean=28.50, SD=1332.285
Request Size [Byte] / Mean=318.59, SD=179.46
Uniform (a=0, b=700) Mean=350, SD=202.07
57 * Assuming MB = MiB (Mebibyte; 230)
HTTP Traffic Model - 2
• With RTT=10ms & synthetic traffic statistics:
– Avg. web page (session) delay: 2.3s
• = RTT*(1+E[Nembedded_object])+E[TParsing]
• Ignoring transmission delay
– Avg. session period (including reading time): 30.8s
– Avg. load (# of bits/session period): 30.4 kbit/s
• = 3803.2 B/sec
• 328670+ sessions needed to fully load 10 Gbit/s line! (328+ for 10 Mbit/s line)
58
Streaming Video Traffic Model • HDTV quality, realistic, high bit-rate video traffic models
are needed for NGOA
– Use H.264/AVC video traces
– “Terminator 2” VBR clip from ASU Video Trace Library • Duration: ~10 min
• Encoder: H.264 FRExt
• Frame Size: HD 1280x720p
• GoP Size: 12
• No. B Frames: 2
• Quantizer: 10
• Mean frame bit rate: 28.6 Mbit/s
– ~334 streams needed to fill 10 Gbit/s line with the following assumption:
» Total overhead: 66 Bytes
• RTP(12), UDP(8), IP(20), and Ethernet (26)
» RTP payload: 1460 Bytes
• In case of 1500-byte MTU 59
FTP Traffic Model - 1
Parameters Probability Distribution Function (PDF)
File Size [Byte] / Mean=2 M, SD=0.722 M, Max=5 M
Truncated lognormal (=14.45, =0.35, max=5 M) Mean=1995616(~2 M), SD=700089.8(~ 0.70M)
Reading Time [sec] / Mean=180
Exponential (l=0.006) Mean=166.667, SD=166.667
Request Size [Byte] / Mean=318.59, SD=179.46
Uniform (a=0, b=700) Mean=350, SD=202.07
60
FTP Traffic Model - 2
• With RTT=10ms & synthetic traffic statistics:
– Avg. session delay: 10 ms
• = RTT
• Ignoring transmission delay
– Avg. session period (including reading time): 166. 68 s
– Avg. load (# of bits/session period): 95.8 kbit/s
• = 11972.7 B/sec
• 104384+ sessions needed to fully load 10 Gbit/s line! (104+ for 10 Mbit/s line)
61
TCP Parameters
• Protocol: Reno
– New Reno, SACK?
• MSS: 1460
– Default: 536
– Optimized for Ethernet interface
• 1452 for PPP
• Advertised window:
– Default: 14*MSS
62
Network Parameters
• Queueing policy: Drop Tail
• Ethernet NIC frame buffer size: 10k frames
– Based on RTT(10ms) * BW(10Gbit/s)
• (10ms * 10 Gbit/s) / 1518*8 bits/frame
• OLT VoQ and ONU FIFO queue sizes: 121,440,000 bits
– Corresponding to 10k maximum size (1518 bytes) Ethernet frames
63
Simulation Runs
• 30 hours with 20-min warmup-period
• 10 replications
– Corresponding to about 3000+ HTTP/500+ FTP sessions
• Rule of thumb (e.g., SMPL text book) recommends at lest 2500 samples with 5 replications.
– 20+ replications for normal distribution assumption in t-test in the future
64
Post-Processing
• Statistical analysis: R
• General scripting: Perl, Python, Shell (bash)
• Plotting: SciPy with matplotlib
65
Initial Results: Elasticity of Hybrid TDM/WDM-PON
• The figure shows the min. number of tuneable transceivers (min(Ntx)) of hybrid PON to achieve the ECR of Rtarget Gbit/s. – For the reference architecture,
the number of transceivers is fixed to the number of ONUs (subscribers), independent of the network load.
– The hybrid PON, on the other hand, can adjust the number of transceivers depending on the network load.
66
V. Summary
67
Summary • We have been investigating the issues of quality of experience (QoE), elasticity*, and
energy efficiency in NGOA with a focus on the solutions beyond 10G-EPON/XG-PON (i.e., NG-PON2 by ITU-T) and found out that the progress in the clean-slate design of NGOA has been impeded by the absence of a comprehensive research framework for a comparative analysis of candidate network architectures and protocols.
• We propose, therefore, a research programme for the clean-slate design of NGOA with the following major objectives:
– A new comparative analysis framework based on statistical hypothesis testing and user-perceived performances;
– User behaviour modelling for future services and applications;
– Virtual test bed for experiments under a realistic operating environment;
– Energy-efficient and elastic components and architectures.
• The results of the proposed research programme will not only enable researchers to carry out a fair and objective comparison of various candidate architectures and protocols, but also allow network operators & service providers to dimension their future NGOA under a realistic environment through the virtual test bed.
68 * The elasticity in the context of networking means the ability to manage overall
performances by fast provisioning of network resources based on user demands.