A Study of Bandwidth-sharing Mechanisms in Connection-oriented Networks Ph.D. Dissertation presented by Xiangfei Zhu Department of Computer Science University.

A Study of Bandwidth-sharing Mechanisms in Connection-oriented Networks

Ph.D. Dissertation presented by

Xiangfei ZhuDepartment of Computer Science

University of VirginiaFeb 19, 2008

2/19/2008

Ph.D. Dissertation Defense Department of Computer Science, University

of Virginia 2

Outline

Quick overview Hypothesis and Metrics Contributions and Publications

Motivation Proposed mechanisms

BA-n BA-First VBDS Immediate-request

Related work Summary

2/19/2008


of Virginia 3

Hypothesis

Well-designed algorithms employing immediate-request and book-ahead bandwidth-sharing mechanisms will lead to efficient utilization of modern high-speed connection-oriented networks

2/19/2008


of Virginia 4

Metrics

Service provider-oriented metrics Utilization Always possible to achieve high utilization if there are no user-oriented

performance requirements

User-oriented metrics Call blocking probability: book-ahead mechanisms for session-type requests Delay: book-ahead mechanisms for data-type requests

Combined metrics Session type: express call blocking probability as a function of utilization Data type: express mean transfer delay as a function of utilization

2/19/2008


of Virginia 5

Key contributions

Two book-ahead mechanisms for session-type requests Analytical and simulation models for these two schemes Models can be used as tools to test design choices and parameter values

A book-ahead mechanism for data-type requests Overcomes a disadvantage of using circuit-switched networks for file transfers

(when compared to packet switching)

Design and deployment of a wide-area, high-speed, optical dynamic circuit network Demonstrated the readiness of off-the-shelf switches for actual service offerings

Measurements of actual end-to-end call setup delays and per-switch processing delays Useful to other researchers for modeling purposes

2/19/2008


of Virginia 6

Publications

X. Zhu and M. Veeraraghavan, " Analysis and Design of Book-ahead Bandwidth-Sharing Mechanisms," accepted by the IEEE Transactions on Communications (TCOM).

X. Zhu, M. E. McGinley, T. Li, and M. Veeraraghavan, "An Analytical Model for a Book-ahead Bandwidth Scheduler," Proc. of IEEE Global Telecommunications Conference (Globecom) 2007, Washington, DC, Nov. 2007.

X. Zhu, X. Zheng, and M. Veeraraghava, "Experiences in implementing an experimental wide-area GMPLS network," IEEE Journal on Selected Areas in Communications (JSAC), vol. 25, pp. 82-92, Apr. 2007.

X. Zhu, X. Zheng, M. Veeraraghavan, Z. Li, Q. Song, I. Habib, and N. S. V. Rao, “Implementation of a GMPLS-based network with end host initiated signaling,” in Proc. Of IEEE International Conference on Communications (ICC) 2006, Istanbul, Turkey, Jun. 2006.

2/19/2008


of Virginia 7

Why the renewed interest in connection-oriented networks? Internet – connectionless packet-switching

Pros: efficient (high utilization) Cons: low quality of service (bandwidth, delay, jitter, etc. )

Resurgence of interests in connection-oriented networks: Top-down driver: large-team scientific projects require

predictable high-speed network services Bottom-up driver: advances in optical circuit-switching

technologies

Various connection-oriented testbeds are being deployed around the world NSF Experimental Infrastructure Network (EIN)

program ESnet4 (US), CA*net4 (Canada), UKLight (UK),

SURFnet (Netherlands), JGN2 (Japan) Internet2 Dynamic Circuit network

Terascale Supernova Initiative (TSI)http://www.phy.ornl.gov/tsi/

Large Hadron Collider (LHC)http://www.phys.ufl.edu/~matchev/LHCJC/lhc.html

2/19/2008


of Virginia 8

Internet2 deployment of Dynamic Circuit network

Backbone picture reprinted from http://www.internet2.edu/pubs/networkmap.pdfIP NetworkDynamic Circuit Network

2/19/2008


of Virginia 9

Why revisit the topic of bandwidth sharing in connection-oriented networks?

Existing mechanisms Immediate-request (IR) mode: used in the telephone network Leased-line mode: used in high-speed connection-oriented networks, such as

SONET and WDM Can these mechanisms be used in connection-oriented networks in new

context (high-speed + new apps)? IR mode: cannot achieve high utilization with low call blocking probability when

channel density is low Channel density in the telephone network is on the order of 100 or more Channel density in high-speed testbeds is on the order of 10

Leased-line mode: poor temporal sharing, expensive and inefficient Cannot be used because the number of universities involved in these projects is large

Better utilizationBetter service quality

Leased linesImmediate

request

New bandwidth-sharing mechanisms are needed!

2/19/2008


of Virginia 10

What mechanisms exist for sharing resources in other contexts?

Reservation systems Reservation phase before resource usage e.g., book flight tickets, make medical appointments, etc.

Queueing systems On-demand service e.g., bank teller, grocery store checkout, etc. Two types of queueing system based on waiting space

Bufferless queueing – no waiting space e.g., street parking

Buffered queueing – has waiting space e.g., bank teller, grocery store checkout

2/19/2008


of Virginia 11

idleidle

Are these mechanisms suitable for bandwidth sharing?

Reservation systems Yes, book-ahead mode

Queueing systems Bufferless queueing – Yes, immediate-request call-blocking mode Buffered queueing – No

X1 X2 X3 H1 H2

H3 H4 H5

H6 H7 H8

2/19/2008


of Virginia 12

Two types of book-ahead systems

Classification based on request type Session-type requests

Specify desired bandwidth and duration e.g., remote visualization and remote instrument control

Data-type requests Specify size of data to be transferred e.g., file transfers

File size known at the sending end

2/19/2008


of Virginia 13

Proposed mechanisms

Bandwidth Sharing in high-speed connection-oriented networks

Immediate-requestBook-ahead

BA-n/BA-First VBDS (Varying-Bandwidth Delayed Start)

Low-to-moderate per-channel ratehigh per-channel rate

session-type requests data-type requests

BA-n BA-First

Users specify a set of call-initiation time options

Users accept any call-initiation time

Deployed a testbed

Implemented software

Measured call-setup delays

Analytical model

Simulation model

Comparison with IR

Analytical model

Simulation model

Comparison with IR

Simulation model

Comparison with packet switching

Published in TCOM Published in Globecom

Published in JSAC

Published in ICC

2/19/2008


of Virginia 14

Channel available for H timeslots starting at any one of the n call-initiation times? Yes, accept request No, reject request

Analytical model for the BA-n scheme

X XSwitch1 Switch2

scheduler

A call specifies: - Bandwidth: 1 channel

- Holding time: H timeslots

- Set of n call-initiation times: {s1, s2,…, sn}

m channels

Assumptions: Call arrival process is Poisson

2/19/2008


of Virginia 15

Discrete-time Markov Chain model

System state: vector X with K components (x1, x2, …xK) xi: number of reserved channels in the ith interval 0≤xi ≤ m

Challenges Non-homogeneous system

Transition rates at time interval boundaries are infinite, but finite at other times Mixed system

Call arrival process: continuous Call holding time: discrete

A user can reserve any timeslots in the reservation window

Key insights Embedded DTMC at time interval boundaries Discretize time into very “small” timeslots to use geometric distribution to approximate (exponential) call

interarrival time distribution Timeslots should be small enough to make the probability of more than 1 call arriving in a timeslot negligible Any call arrival rate can be downgraded to a small call arrival rate by changing the time unit

e.g., 36 call/hour -> 0.01 call/second

(x1, x2, …xK)m: link capacity in channels

K: reservation window in timeslots

2/19/2008


of Virginia 16

Simulation model

Limitation of the analytical model – does not scale with m Recall that the state space is defined as

Size of the state space: (m+1)K

Simulation model Support larger values of m Relax assumptions used in the analytical model

Call-initiation time options: uniform distribution → bell-shaped distribution

Per-call bandwidth: single channel → multi channels Path length: single link → multi links

2/19/2008


of Virginia 17

Model validation and verification

Model validation Our models are for an initial design and

implementation of BA systems Therefore, no real-world measurements Model validation technique – peer/expert reviews

Real system measurements “available” for input parameters Example:

Real-system measurements for telephony applications - Poisson call arrival process

Same pattern likely in video-conference calls

Model verification Compare analytical model results with simulation model results

[Jain91] R. Jain, The Art of Computer Systems Performance Analysis: Techniques for Experimental Design, Measurement, Simulation, and Modeling, New York, Wiley-Interscience, 1991.[Pace02] D. K. Pace and J. Sheehan, “Subject matter expert (SME)/peer use in m&s v&v,” in Proc. of the Foundations, Lauarel, MD, Oct. 2002.

“Three validation techniques Expert intuition Real system measurements Theoretical results” [Jain91]

“Qualitative validation has to be used when adequate acceptable real world data do not exist to permit quantitative validation and is based mainly on SME (Subject Matter Expert) and peer view” [Pace02]

“Three aspects of model validation Assumptions Input parameter values and distributions Output values and conclusions” [Jain91]

2/19/2008


of Virginia 18

Key results from the BA-n study

BA-n scheme outperforms IR scheme when per-channel rate is high e.g., when m=10

With the IR mode, high utilization achievable but at a cost 23% call blocking probability at 80% utilization 46% call blocking probability at 90% utilization

With the BA-3 mode, high utilization achievable with low call blocking probability 0.1% call blocking probability at 80% utilization 2% call blocking probability at 90% utilization

Reservation window size (K) dependence on call holding time (H) K/H does not need to be large e.g., when m=10, to achieve 90% utilization with 2% call blocking probability,

K=4H.

Multi-link scenario BA-n scheme outperforms IR Fairness achieved with “trunk reservation”

Between long-path and short-path calls

2/19/2008


of Virginia 19

Roadmap



BA-n/BA-First VBDS

high per-channel rate


BA-n BA-First

calls specify a set of call-initiation time options

calls accept any call-initiation time

Analytical model

Simulation model

Comparison with IR

Analytical model

Simulation model

Comparison with IR

Simulation model



Published in JSAC

Published in ICC

Deployed a testbed



Low-to-moderate per-channel rate

2/19/2008


of Virginia 20

m channels

Is a channel available in the entire reservation window? Yes, accept request No, reject request

Analytical model for the BA-First scheme

X XSwitch1 Switch2

scheduler

A call specifies: - Bandwidth: 1 channel

- Holding time: H timeslots

- Set of n call-initiation times: {s1, s2,…, sn}

- Any call-initiation time

1 timeslot

Assumptions: Call arrival process is Poisson

2/19/2008


of Virginia 21

System state

Use “bins” to represent reservation intervals If the ith bin is not full, all bins after it must be empty

The system state is expressed as a 2-tuple (i, n) i – index of the first bin that is not full n – number of reserved channels in the ith bin A special case is (K, m), which denotes the state in which all bins are full

Call arrivals

m

n

2/19/2008


of Virginia 22

m

n

The state of the system changes in two cases A call arrives:

e.g., (i, n)->(i, n+1) if n<m-1; (i, n)->(i+1, 0) if n=m-1 and i<K A time-interval boundary is encountered

e.g., (i, n)->(i-1, n) if i>1 The model is a CTMC but it is non-homogeneous

The system behavior at the timer-interval boundaries is different from its behavior at other times

There is an embedded time-homogeneous DTMC if we only look at the system at the time-interval boundaries

Call arrivals

Call arrivals

CTMC

2/19/2008


of Virginia 23

m

n

The transition probability can be calculated by counting the number of calls (denoted by a) that arrived in the past time interval, and calculating the probability that a calls arrive in a interval

A: number of call arrivals in the current interval FA(a) is the Cumulative Distribution Function of A

GA(a) is the Probability Mass Function of A

The transition probability from state (i, n) to state (j, q) is 1-FA(mK-1) if i=1 & (j,q)=(K,m), i.e., mK or more calls arrived

GA(m(j-1)+q) if i=1 & (j,q)≠(K,m), i.e., m(j-1)+q calls arrived

1-FA(m(K-i+1)+m-n-1) if i≠1 & (j,q)=(K,m), i.e., m(K-i+1)+m-n or more calls arrived

GA(m(j-i+1)+q-n) if i≠1 & (j,q)≠(K,m), i.e., m(j-i+1)+q-n calls arrived

j

q

Call arrivals

Call arrivals

Embedded DTMC

2/19/2008


of Virginia 24

Performance metrics

Call blocking probability

Link utilization

Mean scheduling delay - two parts Integral part: number of intervals before scheduled service interval Fractional part: delay within the arrival interval

Call arrivals

integral partfractional part

m

n

2/19/2008


of Virginia 25

IR v.s. BA schemes

Example To achieve a 90% utilization

with a call blocking probability less than 10% BA-First schemes are needed

when m<59

To achieve a 90% utilization with a call blocking probabilityless than 20% BA-First schemes are needed

when m<32

Use of the model -

Test design choices and parameter values

2/19/2008


of Virginia 26

Parameters: 1) link capacity in channels, m = 2 or 8 2) reservation window size, K = 2, 4, 8, or 16

To run a system at 100% offered load with a 4% or less call blocking probability If m=2, K should be 8 time units If m=8, the number is only 4 time units

Is larger value of K always better? If m=8, call blocking probability and utilization plots for K=4, 8 and 16 overlap But mean scheduling delay increases significantly as K increases

Use of the model -

Select an appropriate reservation window size

Increasing reservation window size beyond a certain level is actually detrimental to system performance!

2/19/2008


of Virginia 27

Use of the model -

An approximate solution for M/D/m/p system

Solutions exist for M/D/1, M/D/m (approximation) systems No existing solution for M/D/m/p system BA-First model (m, K) ≈ M/D/m/m(K+1) queuing model at moderate-to-high loads

Why? Call-arrival process: both Poisson Call holding time: both deterministic Reservation window is effectively “waiting space”

1/2

fractional part

2/19/2008


of Virginia 28

We modeled the BA-First mechanism using a non-homogeneous CTMC

We extracted an embedded DTMC and solved it for steady-state probabilities

We obtained solutions for metrics such as call blocking probability, link utilization, and mean scheduling delay

We demonstrated the use of the model as a design tool for book-ahead systems

We demonstrated the use of the model as a solution for M/D/m/p queueing system at moderate-to-high loads

Key results from the BA-First study

2/19/2008


of Virginia 29

Roadmap



BA-n/BA-First VBDS



BA-n BA-First



Analytical model

Simulation model

Comparison with IR

Analytical model

Simulation model

Comparison with IR

Simulation model



Published in JSAC

Published in ICC

Deployed a testbed




2/19/2008


of Virginia 30

Book-ahead scheme for data-type requests

Data-type requests: specify size of data to be transferred

Drawback of using circuits for file transfers With fixed-bandwidth allocation, file transfers cannot take advantage of bandwidth

freed by the completion of other transfers

Fixed-Bandwidth Delayed Start (FBDS) Fixed-bandwidth allocation with rate set to maximum rate

Varying-Bandwidth Delayed Start (VBDS) Assign different bandwidth levels for different time ranges

File transfer request =(File Size, Maximum rate, [Requested start time])

Can be provided by file server

Limited by various constraints at end hosts, such as disk-access speed

2/19/2008


of Virginia 31

VBDS

Idea of VBDS Upon receiving a reservation request, VBDS scheduler returns a Time-

Range-Channel (TRC) vector {(Bk, Ek, Ck, k=1,2,…)} Bk: start time of the kth time range

Ek: end time of the kth time range

Ck: set of channels allocated to the transfer in the kth time range

Scheduler maintains channel-availability function γ(t)

Cost of VBDS Switches need to be reprogrammed multiple times within a transfer

Switch programming time is considered in the analysis Switches need to maintain channel availability function

Reduce the number of changes in channel-availability function Discretize time

2/19/2008


of Virginia 32

VBDS example

Assumptions: 4-channel link with per-channel rate 10Gbps Unit of time discretization: 100ms Switch programming time: 1 unit

A file transfer request specifies (5GB, 20Gbps, 50) (50, 60, {4}) – 1.125GB (60, 70, {2, 4}) – 2.375GB (70, 75, {2, 4}) – 1.5GB

(File Size, Maximum rate, Requested start time)

0 10 20 30 40 50 60 70 80 … ∞

Ch

an

ne

l ava

ilab

ility

γ(t

)4

3

2

1

Time

2/19/2008


of Virginia 33

Numeric results

Compare VBDS, FBDS, and Packet switching (PS)

Nor

mal

ized

Del

ayA

vera

ge th

roug

hput

2/19/2008


of Virginia 34

Key results from the VBDS study

Circuit-switched network with VBDS achieves similar performance as packet-switched networks for moderate-to-large files Significant: at high speeds, circuit switching cost << packet switching cost

VBDS favors large files when compared to packet switching Packet switching: newly arriving transfers “cut in” VBDS: Not so. Allocated bandwidth remains dedicated to ongoing transfers

We do not recommend using circuit-switched network for small files Scheduling and circuit setup overheads

Cost Circuit switching: setup overhead (unsuitable for small files) Packet switching: congestion control algorithm (lower throughput for moderate-to-large

files)

2/19/2008


of Virginia 35

Roadmap



BA-n/BA-First VBDS



BA-n BA-First



Analytical model

Simulation model

Comparison with IR

Analytical model

Simulation model

Comparison with IR

Simulation model



Published in JSAC

Published in ICC

Deployed a testbed




2/19/2008


of Virginia 36

Immediate-request bandwidth sharing

Deployed a wide-area experimental network with immediate-request mode of bandwidth sharing - CHEETAH

State-of-the-art in 2004 Control-plane protocols are standardized by IETF - GMPLS

protocol suite Vendors have implemented these protocols in high-speed optical

circuit switches No deployed network uses these functions No signaling protocol client for end hosts to enable the creation

of end-to-end high-speed circuits

2/19/2008


of Virginia 37

CHEETAH network

Switches: Sycamore SN16000 Intelligent Optical switch Robust implementation of GMPLS control-plane protocols Support standardized Ethernet-SONET mapping

Atlanta, GA

zelda1

zelda2

zelda3

Raleigh, NC

OC192card

ControlCard

GbE/10GbE

card

SN16000

H wukong

Oak Ridge, TN

To Cray X1

zelda4

zelda5

HH

OC192card

ControlCard

GbE/10GbE

card

SN16000

H

HH

OC192card

ControlCard

GbE/10GbE

card

SN16000OC-192 OC-192

End hosts: general-purpose Linux PCs with two NICs and CHEETAH end-host software

2/19/2008


of Virginia 38

IR mode of sharing in CHEETAH

Designed and implemented an end-host software package based on the GMPLS architecture Stand-alone circuit request tools Integrated into applications such as Squid (an open-source web proxy

software)

Ran experiments of IR mode call setups and releases

Measured end-to-end circuit setup delays and per-switch signaling message processing delays Measurements useful to other researchers for modeling purposes

Demonstrated the readiness of off-the-shelf switches for actual service offerings

2/19/2008


of Virginia 39

Related work

Research papers on book-ahead bandwidth sharing Most of these papers use simulations None of them considers book-ahead calls with multiple acceptable options Our results show that a book-ahead mechanism that specifies only one call-

initiation time may perform worse than an immediate-request mechanism

File transfers List scheduling: all proposed algorithms use fixed allocations Bin packing: cannot break a block into pieces to fit into bins TCP improvements: determine fair share for a flow faster and more

accurately, while we determine share for a flow during setup

Optical connection-oriented testbeds e.g.: ESnet4, NSF DRAGON, CA*net4, UKLight, JGN2, etc. Focus: implementation & inter-domain usage Our work: mixed study of IR and BA; theoretical modeling + implementation

2/19/2008


of Virginia 40

Summary

High-speed connection-oriented networks should support a combination of bandwidth-sharing services

Greater sharingBetter service quality

IP servicesLeased

linesImmediate

request

BA-n/BA-First VBDS

Existing services:

New services:

For video telephony, transfers of moderate-sized files

For reservations that specify file size (large file transfers)

For reservations that specify desired bandwidth and duration

For serving as “wires” between switches to create networks that offer other services

2/19/2008


of Virginia 41

Future work

Routing issue in reservation phase Currently assume a linear topology in multi-link scenarios Multiple route options should be exploited

Distributed implementation Necessary for inter-domain scheduling

Service providers do not share network topology information with each other

Validate models against real measurements A long-term future work item after deployment & user base build-up

Questions from Form G111

2/19/2008


of Virginia 43

Questions from Form G111 -

Defining the problem

In the context of new optical circuit-switched technologies and new application requirements, what bandwidth-sharing mechanisms can lead to efficient utilization of modern high-speed connection-oriented networks?

2/19/2008


of Virginia 44


Analysis of previous and related work

Research papers on book-ahead bandwidth sharing Most of these papers use simulations None of them considers book-ahead calls with multiple acceptable options Our results show that a book-ahead mechanism that specifies only one call-

initiation time may perform worse than an immediate-request mechanism

File transfers List scheduling: all proposed algorithms use fixed allocations Bin packing: cannot break a block into pieces to fit into bins TCP improvements: determine fair share for a flow faster and more accurately,

while we determine share for a flow during setup

Optical connection-oriented testbeds e.g.: ESnet4, NSF DRAGON, CA*net4, UKLight, JGN2, etc. Focus: implementation & inter-domain usage Our work: mixed study of IR and BA; theoretical modeling + implementation

2/19/2008


of Virginia 45


Success criteria

Has the student adequately defined the measure(s) of success to be used to evaluate the work? Is there a well defined metric with a goal? Does the metric adequately represent the desired success criteria?

Success criteria Session-type BA requests

BA-n: better performance than IR BA-First: a model that scales to m>100

Data-type BA requests At least the same performance as packet switching

IR mode Stable network deployment and software implementation

Metrics Session type: express call blocking probability as a function of utilization Data type: express mean transfer delay as a function of utilization

2/19/2008


of Virginia 46


Solution

Is the approach taken well executed? Does it appear to be correct? Is the work technically challenging? Does the student utilize appropriate professional standards?

A combination of analytical, simulation, and experimental methods. Two book-ahead mechanisms for session-type requests

Analytical and simulation models for these mechanisms One book-ahead mechanism for data-type requests

A simulation model for this mechanism A wide-area testbed for experimental study of the immediate-request

mechanism Testbed deployment Software implementation Experimentation and measurements

2/19/2008


of Virginia 47


Innovation and risk

Two new Markov chain models for book-ahead bandwidth-sharing schemes (first Markov chain models for book-ahead schemes)

An approximate solution for the M/D/m/p queueing system

One of the first deployments of a wide-area high-speed dynamic circuit network

2/19/2008


of Virginia 48


Broader implications(Social, economic, political, technical, ethical, business, etc.)

Demonstrated the readiness of off-the-shelf circuit switches for actual service offering (business and technical)

Designed efficient bandwidth-sharing algorithms for high-speed connection-oriented networks Circuit switches are less complex than packet switches, which

means Less expensive (economic) Consume less power (environmental)

Backup slides

2/19/2008


of Virginia 50

Assumptions Link capacity m = 1 Advance-reservation horizon K = 3 Number of classes L = 2 Holding time for class-1 calls h1 = 1 Holding time for class-2 calls h2 = 2 Number of options n = 1

System transition happens at the end of each timeslot Example: state (0, 0, 1)

A call arrives and reserves the third timeslot -> state (0, 1, 1) Pr=(1/3)pr1 A call arrives and reserves the first timeslot -> state (1, 1, 0) Pr=(1/3)pr1 No call arrives or the arrived call is blocked -> state(0, 1, 0) Pr=1-(2/3)pr1

Current time t t+k

1

# o

f rese

rved

ch

an

nels

Time

(Backup slides) BA-n - Example of the analytical model for the BA-n scheme

2/19/2008


of Virginia 51

Define a left shift operator : If , .

Define a K-component vector , where

The transition probability from state x to state y is

p: the probability that a call arrives during a time slot rj: the probability that an incoming call belongs to class j qi,j: the probability that a class-j call is admitted with a initiation time of the ith timeslot Bx: the probability that an incoming call is blocked when the system is in state x

First row: a class-j call is admitted with an initiation time of the ith timeslot Second row: no call arrives or a call arrives but is blocked Third row: all other states

(Backup slides) BA-n –

DTMC model – Transition probability matrix

2/19/2008


of Virginia 52

Use Hypergeometric probability mass function to calculate Bx and qi,j A large set of N elements, known to have d defective elements The probability of having k defective units in a random batch of n

elements, drawn without replacement from the large set

Define dj: number of “ineligible” timeslots for class-j calls Mapping:

A total of (K+1-hj) candidate timeslots: corresponds to N dj ineligible timeslots: corresponds to d defective units e.g.: the first t options are all rejected: corresponds to a batch of n elements are all

defective (k=n) After we obtain the transition matrix

Calculate the steady-state probabilities Calculate average call blocking probability and utilization


To compute qij and Bx

2/19/2008


of Virginia 53

BA-all clearly outperforms IR BA-1 is worse than IR

Reason: “gaps” are caused by advance reservations Analogy: if a doctor spends exactly 1 hour with each patient, patients arriving in the middle of an hour will

cause gaps (time period shorter than 1 hour) Restricted call-initiation times

Call-initiation time options are restricted to fall on call holding time boundaries Restricted BA-n mechanisms clearly outperform IR Performance of restricted BA-n is almost as good as BA-all

(a) Call blocking probability (b) Utilization


Comparison of blocking probability and utilization

2/19/2008


of Virginia 54

Provide insight into how to select the advance-reservation horizon (K ) Longer K means better performance Longer K also means greater storage and computation needs

The performance improvement is small after K reaches a certain value

(number of class L=1, call holding time H = 200, offered load = 100%)

(Backup slides) BA-n – Dependence of reservation window size on number of channels and call holding time (1)

2/19/2008


of Virginia 55

BA-all with Number of channels m =2 Call holding time H =300 Offered load = 100%

The ratio K/H instead of K determines the call blocking probability. K/H values for different values of m corresponding to 3 values of call blocking probability

Call blocking probability 2% 5% 10%

m=2 14 6 4

m=5 5 4 3

m=10 4 3 2

(Backup slides) BA-n – Dependence of reservation window size on number of channels and call holding time (2)

2/19/2008


of Virginia 56

Consider a system designer who wants to know the payoff by increasing the reservation window size Assume we want to run the system with 1% call blocking probability e.g., m=4, by increasing K from 2 to 4, the system load/channel can be increased from

75% to 93% This is quite significant in that it allows for a 24% increase in the number of endpoints

multiplexed on to the link

m=4

(Backup slides) BA-First –

Numerical results – system design

2/19/2008


of Virginia 57

(Backup slides) VBDS –

Why prefer using connection-oriented networks for file transfers The cost of high-speed circuit switches are lower than high-speed packet switches

High-speed memories for route table lookup and packet buffering Rich set of features such as policing and shaping

The simulated PS system is an idealized system in which buffers are assumed to be infinitely large In reality packet loss will occur due to congestions Mechanisms, such as TCP’s congestion control schemes, are required to recover from

these packet losses with retransmissions and rate adjustments Transport protocols designed for circuits, such as Circuit-TCP (C-TCP) are more efficient

Take advantage of the information on the fair share of a flow Disabling TCP’s Slow Start and AIMD algorithm

Item Cisco 12416 Sycamore SN16000

Base system $130,000 $183,500

10x1GbE card $169,830 $63,500

1x10GbE $125,000 $65,500

1xOC192 $225,000 $37,500

A system with 6 pieces 10xGbE +6 pieces 1x10GbE + 3 pieces OC192

(Total 120Gbps client data rate)

$2,573,970($21,000/Gbps)

$1,084,700($9,000/Gpbs)

2/19/2008


of Virginia 58

(Backup slides) VBDS –

VBDS favors large files when compared to PS

Packet switching: newly arriving transfers “cut in”

VBDS: Not so. Allocated bandwidth remains dedicated to ongoing transfers

2/19/2008


of Virginia 59

(Backup slides) IR -

Erlang-B formula

Cannot achieve high utilization with low call blocking probability when m is small

Call blocking probability (PB) against the link capacity expressed in channels (m)

2/19/2008


of Virginia 60


GMPLS control plane

Purpose of Generalized Multi-Protocol Label Switching (GMPLS) control plane Dynamic bandwidth sharing (distributed) Provisioning (configure the switches for the circuit/VC)

Three components Link management protocol OSPF-TE routing protocol RSVP-TE signaling protocol

Bandwidth sharing mode Immediate-request (cannot specify a future call-initiation time or call

holding time in protocols) Calls are accepted or rejected - “call blocking"

2/19/2008


of Virginia 61


CHEETAH concept

Provide on-demand circuit service as an add-on to the connectionless service provided by the Internet

Hybrid circuits: GbEthernet-SONET-GbEthernet

2/19/2008


of Virginia 62

Architecture

Based on the RSVP-TE code from KOM/DRAGON About 40K lines of C++ code

What I did: Modified the code to inter-operate with the Sycamore SN16000 Added admission control, session management, user interface, etc. Integrated code for DNS lookup from our partner CUNY Designed and implemented APIs for general applications About 4K lines of new code


CHEETAH end-host software development

Application

DNS client

RSVP-TE module

TCP/IP

C-TCP/IPNIC 1

NIC 2

End HostCHEETAH software

Internet

SONET circuit-switched network

CircuitGateway

CircuitGateway

Application

DNS client

RSVP-TE module

TCP/IP

C-TCP/IPNIC 1

NIC 2

End HostCHEETAH software

2/19/2008


of Virginia 63


CHEETAH end-host software architecture

DNS lookup – to support our scalability goal

Five steps of circuit setup Message parsing

RSVPD Route determination

Left to the edge switch CAC Date-plane configuration

Route/ARP table update Message construction

RSVPD

DNS server

RSVP-TE Daemon(RSVPD)

DNS lookupCHEETAH daemon (CD)

socket

CD API socket

User space

Kernel spaceC-TCP

C-TCP API

End host

Circuit-requestor

RSVP-TE

messages

DNS client

CAC

Route/ARP table update

RSVPD API

CHEETAH software

CD API can be integrated into web servers, FTP servers, etc., so that “elephant” flows are automatically handled via a dynamically created dedicated circuit/VC

2/19/2008


of Virginia 64


End-to-end signaling delay measurements

Signaling delays incurred in setting up a circuit between zelda1 (in Atlanta, GA) and wuneng (in Raleigh, NC) across the CHEETAH network.

Observations: Delays for setting up SONET circuits for rates in the original SONET hierarchy are small

(166ms) Delays for hybrid Ethernet-SONET circuits are much higher (1.6s) (vendor implementation)

The measured delay can be used for analytical and simulation models for related research

Circuit type End-tend circuit setup delay (s)

Processing delay for Path message at

the NC SN16000 (s)

Processing delay for Resv message at

the NC SN16000 (s)

OC-1 0.166103 0.091119 0.008689

OC-3 0.165450 0.090852 0.008650

1Gb/s EoS 1.645673 1.566932 0.008697

Round-trip signaling message propagation plus emission delay between GA SN16000 and NC SN16000: 0.025s

A Study of Bandwidth-sharing Mechanisms in Connection-oriented Networks Ph.D. Dissertation presented by Xiangfei Zhu Department of Computer Science University.

Documents

bookahead mechanisms

bookahead bandwidth

design of book

function of utilization

connection oriented

university of virginia2

university of virginia7

sessiontype requests