Simulations K Mills joint work with C Dabrowski NIST December 1 2015 For more details see NIST Technical Note 1901 http wwwnistgovitlantduploadTechNote1901draft1pdf ID: 553282
Download Presentation The PPT/PDF document "The Influence of Realism on Congestion i..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
The Influence of Realism on Congestion in Network Simulations*
K. Mills (joint work with C. Dabrowski)NISTDecember 1, 2015
*For more details see: NIST Technical Note
1901
http
://
www.nist.gov/itl/antd/upload/TechNote1901-draft1.pdf
Slide2
Total talk is 20 slidesMotivation – 6 slidesResearch Questions and Approach – 2 slides
Models – 5 slidesExperiment Design – 1 slideResults – 5 slidesFindings – 1 slide
December 1, 2015
NIST TN 1901
2Slide3
Congestion Spreads across Networks in Space and TimeEmpty 2D Lattice
Congestion Persistent
Congestion Spreading
Congestion Widespread
TIME
IMPLICATION
:
SHOULD BE POSSIBLE TO DETECT THE SPREADING
PROCESS AND PROVIDE EARLY WARNING OF
INCIPIENT CONGESTION COLLAPSE
December 1, 2015
NIST TN 1901
3
PACKET-INJECTION RATE
2D lattice animation taken from
“Percolation Theory
Lecture Notes
”, Dr. Kim Christensen, Imperial College London
, October
9, 2002
Congestion Sporadic Slide4
Spreading Processes often Modeled as Percolation
Below
p
c
no GCC
Above
p
c
GCC forms and grows
to include all nodes
Percolation
spread of some property
in
a
lattice (or graph) leading
to the formation of a
giant connected component
(GCC), as measured by
P
∞
,
the proportion of nodes encompassed by the GCCp is probability a node has property pc is known as the critical point
p < pc no spreadp = pc percolation phase
transitionp > pc spread occurs, and expands with increasing p
Near a critical point, the process exhibits signals,
typically attributable to increasing, systemic correlationNIST TN 1901
4December 1, 2015Slide5
Academics Model Spreading Network Congestion as a Percolation ProcessFinding that Signals Appear
Near a Critical Pointin Abstract Network Models
NIST TN 1901
5
December 1, 2015Slide6
Penn State Researchers (Sarkar, Mukherjee, Srivastav, and
Ray 2009)find increasing transit delay as p > pc = 0.3
Increasing slope
in time
series of selected measured variables could
signal
crossing
a
critical point
, allowing network managers to be alerted prior to
network collapse
Aggregate Avg. Transit Delay (D
*
) vs. Network Load (
λ
)
Sampled Avg. Transit Delays (D) for Four Network Loads (λ)
NIST TN 1901
6
December 1, 2015Slide7
Boston University Researchers (Rykalova, Levitan, and Browe
2010) find increased correlation in time series of packets in transit as p nears pc = 0.2
# Packets fluctuate between 260 and 380
# Packets fluctuate between 18,000 and 34,000
Increasing autocorrelation
in time series could
signal
an approaching
critical
point
, allowing network managers to be warned prior to
network collapse
NIST TN 1901
7
December 1, 2015Slide8
Abstract Models Lack Key Traits of Real Networks
Human-engineered, tiered topologies, with propagationRouter buffer sizes finite
Router
speeds varied
to meet demands, limit losses
Injection from sources and
receivers
only at
lowest tier
Distribution of
sources
and
receivers
non-uniform
Connection of sources/receivers with few varied speeds
Duty cycle of
sources exhibits cyclic behaviorHuman sources exhibit
limited patienceSources transfer flows of various sizes
Flows use the Transmission Control Protocol (
TCP
) to modulate injection rate based on measured congestion
Routers &Links
ComputersUsers
ProtocolsNIST TN 1901
8
December 1, 2015DOES LACK OF REALISM MATTER WHEN SIMULATATING NETWORK CONGESTION?Slide9
Specific Research QuestionsDoes congestion spread in abstract models mirror spread in realistic models?Are some elements of realism essential to capture when modeling network congestion?Are some elements unnecessary?
What measures of congestion can be compared, and how, across diverse network models?NIST TN 1901
9
December 1, 2015Slide10
Research Approach
AbstractModel from Literature
Realistic
Model from Literature
Flexible Network Simulator (
FxNS
)
Basic Model
Behaviors
Factor Into
Realism Elements
that can be
enabled or disabled
Simulate Combinations
of Realism Elements and
Compare Patterns of
Congestion
NIST TN 1901
10
December 1, 2015Slide11
ModelsAbstract EGM Model→high abstraction
Realistic MesoNet Model→high realismFlexible FxNS Model→combinations of realism from low to high
December 1, 2015
NIST TN 1901
11Slide12
The Abstract (EGM) ModelP. E
chenique, J. Gomez-Gardenes, and Y. Moreno, “Dynamics of Jamming Transitions in Complex Networks”, Europhysics Letters, 71, 325 (2005)
Simulations based on 11,174-node scale-free graph,
P
k
~
k
-
γ
&
γ
=2.2, taken from a 2001 snapshot of the Internet Autonomous System (AS) topology collected by the Oregon Router Server (image courtesy Sandy Ressler)
γ
= -2.2
NIST TN 1901
12
December 1, 2015Slide13
Details of the EGM Model
h is a traffic awareness parameter,whose value 0 ... 1.where i is the index of a node’s neighbor,d
i
is minimum #hops
to
destination via
neighbor i, and ci is the queue length of i.
Node Buffer Size
:
∞
for EGM, all packets buffered, no packets dropped
Injection Rate
:
p
packets injected at random nodes (uniform) at each time step
Destination Node
: chose randomly (uniform) for each packetForwarding Rate: 1 packet per node at each time step
Routing Algorithm: If node is destination, remove packet; Otherwise select next-hop as neighboring node i with minimum
diSystem Response: proportion ρ of injected packets queued in the network
Computing
di
Measuring ρ
h = 1 is shortest path (in hops)
A = aggregate number of packetst = time
t = measurement interval sizep = packet inject rate NIST TN 1901
13Slide14
Comparative Simulation Results
p
ρ
h
=1
h
=0.85
p
ρ
EGM Simulations
FxNS
Simulations with
All Realism Elements Disabled
NIST TN 1901
14
December 1, 2015Slide15
The Realistic (MesoNet) Model
K. Mills, E. Schwartz, and J. Yuan, "How to Model a TCP/IP Network using only 20 Parameters", WSC 2010, Dec. 5-8, Baltimore, MD.
Comparisons of
MesoNet
Simulations vs.
FxNS
Simulations (all realism elements
enabled)
for eight
MesoNet
responses are available in
NIST TN 1901 – Appendix A
NIST TN 1901
15
December 1, 2015
Figure 2.Slide16
FxNS Combinations
Dependencies among Realism Elements
34 Valid
FxNS
Combinations
NIST TN 1901
16
December 1, 2015Slide17
Experiment Design
FIXED PARAMETERS
218-Router Topology (Fig. 2)
Routing (SPF propagation delay)
Duration (200,000
ts
per
p
)
VARIABLE PARAMETERS
Packet-Injection Rate
p
(up to 2500)
FxNS
Combination RESPONSESCongestion Spread χ=|Gχ|/|GN
| Connectivity Breakdown α=|Gα|/|GN|
Proportion of Packets Delivered πScaled (0..1) Latency of Delivered Packets δ
NIST TN 1901
17
December 1, 2015
Only concepts in common among all 34
combinations: graph and packetSlide18
Results1,2[1] 136
xy-plots (34 FxNS combinations × 4 responses) are available at: http
://
tinyurl.com/poylful
[2] Related
FxNS simulation data can be explored interactively using a multidimensional visualization created by Phillip Gough of CSIRO: http://tinyurl.com/payglq6Slide19
Results I – Abstract (C0) vs. Realistic (C127)
NIST TN 1901
19
December 1, 2015
Plots for all responses and all 34 combinations available:
http://
tinyurl.com/poylful
C0
C127
C0
C127
C127
C0
C127
C0Slide20
Results II – Congestion Spread χ All Combinations
NIST TN 1901
20
December 1, 2015Slide21
Results III – Connectivity Breakdown α All Combinations
NIST TN 1901
21
December 1, 2015Slide22
Results IV – Packet Delivery π All Combinations
NIST TN 1901
22
December 1, 2015Slide23
Results V – Scaled Packet Latency δ All Combinations
NIST TN 1901
23
December 1, 2015Slide24
FindingsCongestion spreads differently in abstract and realistic models
Hierarchical Router Speeds and TCP very important to modelPacket dropping important to model for accurate packet latenciesPropagation delay not important to model in a continental US network, but would be important to model in topologies where propagation delays exceed queuing delaysCongestion spread, connectivity breakdown and the effectiveness and efficiency of packet delivery can be measured using only two concepts: graphs and packets
NIST TN 1901
24
December 1, 2015