Transcript Document

Teletraffic Lessons for
the Future Internet
Presenter: Moshe Zukerman
ARC Centre for Ultra-Broadband
Information Networks
Electrical and Electronic Engineering Dept.,
The University of Melbourne
Outline
•
•
•
•
•
•
My Research
Background: Evolution, services, network design
optimization, cost and carbon cost, Internet growth,
link utilization, Internet congestion control
Optical Internet model and design options
Example of an optical network performance
analysis problem
Results
Conclusion
My research
•
•
•
•
Queueing theory – bursty traffic – link
dimensioning
Optical network performance and design
Medium access control – protocol
performance analysis and enhancement
Other topics: TCP, Wireless/Mobile
networks
New services - research directions
•
Internet of things (mice)
–
–
–
•
make it work from a traffic point of view
light weight protocols
traffic implications - network dimensioning
HD-IPTV, Virtual reality (elephants)
–
–
–
–
Streaming vs download
network dimensioning
multi-service internet
traffic shaping/policing
• Others, in between, e.g. wideband voice
Moore’s law
and Internet equivalence
• Moore's Law: power and speed of
computers will double every 18-24
months.
• Internet backbone traffic grew from one
Tbit/sec in 1990 to 3,000 Tbit/sec in 1997.
• Number of Internet hosts more than
doubled every year between 1980-2000.
Trend
semiconductor performance
computer performance/$
communications bit/$ before 95
communications bit/$ with DWDM
max. Internet trunk speed in service
Internet traffic growth 69-82
Int. traffic growth 83 (TCP/IP) - 97
Internet traffic growth 97-2000
router/switch max. speed pre 97
router/switch max. speed post 97
Source: L. G. Roberts, Computer, January 2000
Doubling Period
18 months (Moore’s)
21 months (Roberts’)
79 months
12 months
22 months
21 months
9 months
6 months (bubble)
22 months
6 months
World Internet Statistics
World Population: 6,676,120,288
Number of Internet Users 1,407,724,920
Penetration 21.1%
%Growth between 2000-2008 290.0%
Source: www.internetworldstats.com
Design Optimization
Aim: To provide services at
Minimal Cost
Subject to:
Meeting required quality of service
And other practical constraints (including
availability of power)
Google Data Center
The Dalles, Oregon
Source: LA Times (14-6-2006) By JOHN MARKOFF and SAUL HANSELL
“Hiding in Plain Sight, Google Seeks More
Power”
Competing with Microsoft on dominance
but the practical constraint is power
Power consumption ~200 MW (RS Tucker)
Google Data Centre (cont.)
Source: www.techbanyan.com/archives/140
Network Power Distribution
•Switching and Routing 34%
•Regeneration 27%
•Processing 22%
•Storage 10%
•Transport 7%
Reference:
“Data Centers Network Power Density Challenges”
By Alex Vukovic, Ph.D., P.Eng. ASHRAE Journal, (Vol. 47, No. 4, April 2005).
Internet Power Usage
TOTAL Population: 6,676,120,288
Number of Internet Users 1,407,724,920
Penetration 21.1 %
%Growth between 2000-2008 290.0 %
Source: www.internetworldstats.com
Internet Power Usage (cont.)
Today Internet (excluding PCs, customers
equipment, mobile terminals etc.)
uses ~1% of total world electricity usage.
If 2 Billion people have broadband access
(1Mb/s) then ~5%.
If 2 Billion people have broadband access
(10 Mb/s) then ~50%.
Source: R.S Tucker, “A Green Internet”
May 2007, CUBIN Seminar, The University of Melbourne
Design Optimization
Aim: To provide services at
Minimal Cost (do not forget to consider also
direct energy $ + indirect carbon $)
Subject to:
Meeting required quality of service
And other practical constraints
(including availability of power)
The other aspect is utilization
(traditional teletraffic concept)
Link Utilization
Utilization =
Proportion of time the link is Busy.
measure for system efficiency and profit for
telecom providers.
The traditional teletraffic aim has been to
maximize utilization subject to
meeting queuing delay (and loss) requirements.
It’s all about using the scraps!
time
Bursty traffic = low utilization and bad service
time
Smooth traffic = high utilization and good service
A Simple model
X
C
time
P(X > C) < Quality measure
frequency
Bursty traffic
many standard deviations
E[X] = 150 Mbit/sec
C = 1000 Mbit/sec
bit rate
frequency
Smooth traffic
Gaussian
many sources
Bit rate
Chebyshev’s Inequality
E[X] = 850 Mbit/sec
C = 1000 Mbit/sec
P(|X-E[X]| > S) ≤ Var(X)/S2
Internet end-to-end protocols
Transmission Control Protocol
(TCP) – Non-Real Time Traffic
User Datagram Protocol (UDP)
for Real-Time Traffic
Towards All-Optical Internet
“Old” Electronic Internet:
Capacity expensive, buffering cheap
Introduction of DWDM makes capacity cheap
Electronic Bottleneck: O-E-O
but maybe the bottleneck is not this E but the
other one (Energy, or P = Power)
Future All-Optical Internet (?):
Link capacity plentiful, buffering painful (cost,
power, space) and also wavelength
conversion (espacially for OPS) is costly
An Internet Model
Access
Optical Core
Bufferless Optical
Burst/Packet Switching
•
•
•
•
•
Packet Switching but without buffers;
Packets cannot be delayed along the way.
Delay is possible at the edges.
Some multiplexing is possible.
Between packet switching and circuit
switching.
• How efficient can it be?
Optical switch
trunk
trunk
Optical switch without buffers and
without wavelengths conversion
trunk
links
trunk
Trunks and Links
A trunk can be composed of 10 cables
Each cable comprises 100 wavelengths
So a trunk will have 1000 links
Let us focus on one output trunk
Markov chain analysis is a common
approach to evaluate loss probability
Models - no buffers many Pipes
M/M/k/k
Arrival
process
Service Number of
distribution
servers
Buffer places
including at servers
M / M / infinity
A = arrival rate (λ) / service rate (µ)
A = arrivals per service time
M/M/k/k was developed for telephony
“We are sorry; all circuits are busy now; will you
try your call again later”.
Old message from a local exchange of:
Erlang B Formula gives the the probability
that a call is blocked under the M/M/k/k
model.
Recursion for Erlang B Formula:
E0(A)=
1
Blocking probability for traffic A and
n channels
Multiplexing Benefit
A
10,000
1,000
100
10
k
10272
1100
137
24
% Utilization
0.97
0.91
0.73
0.42
Target Blocking probability = 0.0001
With and Without
wavelength conversion
If a trunk is composed of 10 cables and
each cable comprises 100 wavelengths
so a trunk has 1000 links
With wavelength conversion, the bottleneck trunk
has 1000 links (achieves 91% Utilization).
Without wavelength conversion it is divided into
100 mutually exclusive sets each of a particular
wavelength that has 10 links (22% Utilization).
Why if larger A increases utilization?
If the number of busy servers (Q) in an M/M/k/k system is
almost always less than total number of output links k, the
M/M/k/k behaves (almost) like M/M/infinity.
For M/M/infinity, Q is Poisson distributed with parameter A.
Thus, E[Q] = Var [Q] = A.
Poisson => Normal as A (and k) increase.
So
σ[Q]/ E[Q] => 0 as A increases.
The spare capacity (k-E[Q]) , e.g. 5σ[Q], becomes negligible
relative to E[Q] (Recall E[Q] =A).
This is similar to what we saw before.
As A increases we go from:
frequency
Spare capacity
1000 Mbit/sec
150 Mbit/sec
Bit rate
Bursty traffic
frequency
To:
850 Mbit/sec
Bit rate
Smooth traffic
1000 Mbit/sec
M/M/k/k modeling of
OPS/OBS over WDM
Blocking probability is obtained
by the Erlang B Formula
wavelength 3
wavelength 2
wavelength 1
Time
Extensions and technology choices
• Limited number of input links.
=> Engset Model - Still telephony(1918)
• Frozen time when a packet is dumped.
=> Generalized Engset Model (Cohen 1957)
• Optical buffers.
• Frozen time - packet is inserted into the buffer.
• Hybrid circuit/packet switching.
• Hybrid electronic/optical switching(!)
• Optical burst switching
• Network with multiple bottlenecks.
• TCP on top.
One optical network model
Core switches:
symmetrical
Edge routers:
infinite buffers;
Access links:
smaller
bandwidth than
core links;
TCP sources:
saturated; no
maximum
window limit;
(conservative,
large send and
receive buffers)
Focus on one output trunk
Notation
M: total number of input links,
K: number of output links,
B: buffer size,
: service rate of a single output link
= reciprocal of mean packet time.
PD: packet loss probability.
Analytical model
λI
PD
λI = 1/[inter packet time per link]
Model of TCP throughput
Relationship between TCP bottleneck
throughput and packet loss probability:
Ragg
N 1.5 / PD
MI


RTTH
(I   )
Ragg : the aggregate TCP throughput,
N : the number of TCP flows,
M : the number of input links,
RTTH : the harmonic average round-trip
time
Generalized Engset with Buffer (GEB)
* = 1/(1/ I +PD/), fixed-point solution is needed.
Related models
• Engset with buffer (EB)
– Use I instead of * in GEB (no need for a
fixed point solution).
• M/M/K/K+B
Fix-point equations
binary search algorithm => fixed point solution
Open loop
Model Validation
16 input trunks
Zero Buffer – Scaling Effect
No wavelength conversion
# Sources
Conclusion
Teletraffic models can be used to provide insight
into the economics of the optical-Internet.
Power usage and related cost must be considered.
 In the optical Internet buffering can be pushed to
the edges efficiently as traffic, number of sources
and capacity (number of wavelengths per cable)
increases, if cost effective optical wavelength
conversion is available.