Professor Michael Devetsikiotis gave a lecture on "Networked 3-D Virtual Collaboration in Science and Education: Towards 'Web 3.0' (A Modeling Perspective) " in the Distinguished Lecturer Series - Leon The Mathematician.
More Information available at:
http://goo.gl/U5nGq
Networked 3-D Virtual Collaboration in Science and Education: Towards 'Web 3.0' (A Modeling Perspective)
1. Networked 3-D Virtual Collaboration in Science
and Education: Towards ‘Web 3.0’
(A Modeling Perspective)
Michael Devetsikiotis
Professor of Electrical & Computer Engineering
North Carolina State University
mdevets@ncsu.edu
http://www4.ncsu.edu/~mdevets
Collaborators: Mitzi Montoya, George Michailidis
NC State Team: Michael Kallitsis, Vineet Kulkarni,
Ioannis Papapanagiotou, Nilesh Gavaskar, Yan Wang
2. Strategic Trends and Overview
• Main themes:
– Services dominate (“service oriented networks”?)
– Ubiquitous social nets and virtual collaboration
– Distributed and virtualized delivery (the “cloud”)
– Convergence: telecom - services - infrastructure
• Horizontal integration and fixed-to-wireless convergence
• NGN, IMS/SIP, web services: middleware meets the telcos
• New apps: p2p, virtual worlds, social nets, games,
virtual collaboration, tele-presence, Web 2.0,…
• Emerging “application content infrastructure”
• All via next generation communication networks
3. Overview: Service Oriented Networks
• Our larger goals:
– Capture “presence” and location awareness
– Quantify socio-technical interactions
– Characterize workload spatially and temporally
– Optimize over multiple resources, across layers
SON and
Convergent Nets
“Cloud”
Social Apps and Workload Aggregation
Virtual and Distributed
Collaboration Delivery
4. Research Topics
• Service Oriented Networking
• Resource optimization in net appliances
• Virtual collaboration environments and socio-
technical modeling
• Resource optimization in clouds and wireless
• Aggregation architectures and traffic models
5. Services in Networks and Economy
• Over 70% of advanced economies today in services
• Components becoming “commodities”
• Applies to telecom and IT sectors too
• Services are about “co-production” and “innovation”
• A new “Service Sciences” discipline is emerging
• Both human level and software/middleware
Business/Economics
Competition
Services/Innovation Technology/Resources
Flexibility Congestion, QoS
6. Overview of Service-Oriented Networking
• Definition
– Service-Oriented Networking (SON): emerging
network architecture gaining IT efficiency by providing
intelligent functionality in the network fabric, previously
unavailable or impractical to implement.
• Details
– Application awareness in the network fabric is key
– Challenges end-to-end principle of networks (“don’t
touch the payload”)
– Assumes that the network can make intelligent
decisions based on application data
– Revisits earlier research in application-aware networks
– NGN standards make architecture more flexible
9. Service-Oriented Networking Functions:
Functional Offloading
• Offload services into the network
fabric that can leverage
specialized hardware
(cryptographic or XML
processing ASIC/FPGA)
• In this example, the network
offers a value added service of
securing SOAP/XML requests
and responses inline
• In certain situations, the network
could provide a full offload of
endpoint services (e.g., caching
stock prices), and would be
managed by a caching policy
10. Service-Oriented Networking Functions:
Content-Based Routing
• Content-based routing typically
involves applying a rule against
some part of a service request
(header or content) to derive a
token as a result.
• This token is then used to make
a routing decision
• In this example, where requests
are XML messages, we utilize
XPath to extract the appropriate
routing token
• This value-added service can be
used to enable service
partitioning (higher efficiency)
12. Challenges in building Service Oriented
Networks
– Scalability of the network with network entities
– Adaptation of network to changes in state
– Distributed policy-driven dissemination of network
management data between nodes
– Distributed control of the network to connect
consumers and providers while enforcing appropriate
policies
14. Our Research in Network Services
• Service networking and optimization
– Service delivery pricing and optimization
– Service-oriented networking
– Architecture for service brokering and delivery
– Measurement-based control of service centers or
“appliances”
– Virtualized server characterization and control
• Networked virtual collaboration
• Cross layer and wireless design
– WiFi and WiMax QoS modeling
– Cross layer modeling, simulation, optimization
– Mesh and multihop systems (WiFi, WiMax,
hybrid)
26. Distributed MBORA: Evaluation
• Used deterministic network calculus for end-to-end delay
• Recently used stochastic calculus providing tighter bound
• Approximate -- need better formula to include processing
delays
• Gauss-Seidel versus Dual Decomposition
• Working on better understanding and more alternatives
28. Cloud computing & Virtual Collaboration
Enterprises are moving towards the application of
Virtual Worlds for internal deployment
Virtual Worlds enable ubiquitous presence and
virtual collaboration
Apply same paradigm in education:
Access applications via a virtual world Synergistic work
and parallelization
Student 1: MatLab; Student 2: OPNET and vice-versa
DE students will interact with their colleagues
No commute needed for students working in industry
29. VIRTUAL COMPUTING LAB (VCL)
“The Virtual Computing Lab (VCL) is a remote access service
that allows to reserve a computer with a desired set of applications
for yourself, and remotely access it over the Internet”
• Users have remote desktop
access to machines loaded,
on demand, with the desired
software.
• Anytime-anywhere access
to applications, transparent to
users.
• Ease of system
configuration and
management, and scalability.
Does not support collaboration among users, yet!
30. VCL 3.0: a motivating example
• Users request their applications from VCL
• An image of a virtual world with those apps is created
• Remote connections are created to those apps from inside the world
• Resources to virtual machine are given according to socio-technical
characteristics of the group members
31. Allocating cloud resources
• Which virtual
machines should
be placed for
execution?
• How do we
optimally
allocate cloud
resources?
35. Social Distance – Connectivity Graph
• Construction of connectivity graph
– Bandwidth availability of each user
– Physical distance (implies communication delay)
– Business distance
– Group size, level of trust between collaborators, etc could also be used
• Use of graph’s diameter to differentiate between different connectivity graphs
• Main idea: assign more resources to group with smaller social-distance
• Larger social-distance: conditions not favoring high collaboration quality
38. Measuring Virtual Presence
Analyze social networking patterns in relation to
communication patterns
Measure degree to which VW creates a sense of virtual
presence in the user
Assess collaboration quality in VWs
Evaluate network traffic in relation to social networking
and communication patterns (e.g., chat communication,
voice, video)
41. Measuring Virtual Presence
(Wonderland/Matlab)
A group-based collaboration exercise in Wonderland
Matlab exercise
Data collected from 120 students (Spring 2010, ECE 220)
Measures include:
Average time per solution
Individual contribution to team performance
Perception of virtual presence, communication mode (voice/chat),
collaboration quality
46. Measuring Virtual Presence
(Opensim/Maze)
A negotiation and planning game in OpenSim
LOST (in 3D maze)
The game requires players to collaborate (lead/follow) in
order to meet a common objective under time constraints
47. Measuring Virtual Presence (cont’d)
• Repeat experiment with different levels of:
– CPU (using cpulimit command)
– Memory (using ulimit command)
– Bandwidth shaping (using the trickle tool)
• Obtain objective functions for our optimization problem
• Define the weights for the edges of our connectivity graph
48. Trials with Programmable Bots
• Use programmable bots as subjects for maze traversal
– A guide bot instructs the lost bot how to “escape” the maze
• Vary the amount of server's communication bandwidth
– Introduce dummy bots who chat with each other
– Limit allocated bandwidth using Linux tools
• Vary the concurrent in-world participants
• The metrics which we can capture:
– Frames per Second reported by server
– Maze completion time
• Bots communicate using IM chatting
• Future extension: emulate voice communication
• (Server specifications: 2.8Ghz, 16 core, 64 bit machine, 16 GB
RAM)
49. Bot Results (IM communication)
250 50
45
Frames Per Second (FPS)
• Time completion vs. 200 40
Traversal Time (secs)
35
Concurrent in-world 150 30
25
users 100 20
15
50 Time (secs) 10
• Frames per second vs. 0
FPS 5
0
Concurrent in-world 25 37 50 60
Concurrent Users
75 82 90
users Trickle bandw idth
180
160
140
120
Traversal Time
100
• Time completion vs. 80
60
Trickle bandw idth
Available bandwidth 40
20
0
50 100 150 200
Available Bandw idth
50. Contributions so far
• Optimal and dynamic allocation of cloud
resources (e.g., CPU, memory, network
bandwidth)
• Why consider presence status of users
– Going towards social awareness
– Introducing social distance and connectivity graphs
• Applications: cloud computing & virtual worlds
– NCSU’s VCL
– Amazon EC2
51. Summary
Social-aware optimization framework
Motivation: Resource allocation of cloud resources to
virtual machines that host virtual collaboration environments
User's presence perception needs to be correlated with
tangible resources (CPU, memory, bandwidth)
Future work: Continue trials and experiments to:
Find suitable utility functions per resource
Investigate other important parameters to be used in the
graph weight function
52. Next Steps & Future Plans - I
Model patterns/bundles as service-oriented network for
deployment in CloudBurst (IBM DataPower appliance)
•Analyze network traffic, CPU patterns (also, power
consumption?)
•Obtain the resource requirements of virtual images according to type of
application used and participant social/business type
•Use above information in the virtual machines placement problem
Continue to collect scaling data from bots
Simulation (demo)
Measure maze completion time
Measure Frames Per Second
Change # of concurrent users
Change CPU/memory/bandwidth
53. Next Steps & Future Plans - II
• Placement problem
– Add the green dimension: place virtual machines to also
account for their power consumption
– Use physical space, cooling and power constraints
• Smart Grid extension? Energy appliances?
54. Network-Enabled Collaboration for Innovation
Enabling VIRTUAL ORGANIZATIONS
Mechanisms Partners:
Open Source S/W (Jazz, VCL)
Virtual Public Schools (K-12)
COLLABORATION
Medical Technology
SSME Community
Serious Gaming
PROTOCOLS
Art City
VIRTUAL WORLDS
AND 3Di
NETWORK &
MIDDLEWARE
Social Innovation Industry/University
Commercial Innovation
Virtual
Proximity: Centennial Living Labs
Testing &
Implementation
Virtual RTP
55. ArtCity: Network-Enabled Art
• Autonomic service delivery platform for the Arts
• Enabling artistic virtual organizations and remote interactions by use
of high speed networking and on-demand service delivery.
• Combine network services with virtual collaboration research, and
with hands-on, “living lab” setting on campus (immersive Art Village
in dorm, Centennial trials and pilot event in EBII).
• Use Centaur lab as hub for connectivity.
• RENCI and other telepresence and mixed reality facilities (e.g.,
Cisco)
• Use-cases: wireless-based mobile gaming and virtualized dance
activities: also serve as sources of system performance and
workload measurements and analysis.
• Measurement phase followed by a design phase, where the
algorithms and protocols in Nortel-sponsored wireless mesh trial can
be adapted for optimized performance in real-life setting.
• Our work on service-delivery platforms and resource allocation will
be tested and tried in this environment and its performance will be
tuned accordingly.
57. Nortel SIP and Next Gen Services over Mesh Wireless
• Partnering with Nortel, Carleton University in
Ottawa, Canada, and Cisco
• Analysis: Cross-layer modeling of performance
• Trial: Wireless mesh testbed in EB-II
• Benefits from Centennial campus wireless network
• Emphasize location, distances and “aware”
network
• Building Wi-Fi positioning system in EB-II
• Stage serious gaming trials
58. Social Distance Aware Utility Functions
• Motivation
– Utility Functions defined almost always at the transport
layer
– Social distance of a user to her peers affects desired
utility
• Approach
– Formalize the type of distances (social, effective)
between related entities in a social graph
– Define and solve the Social Distance Aware Resource
Allocation Problem.
59. Social Distance Aware Resource Allocation
• Network is explicitly made aware of the resource requirements.
• Resource allocation decisions happen in terms of parameter
tuning at corresponding protocol layers.
• Better resource allocation decisions possible due to social
context awareness.
62. Resource Allocation in a WLAN
• Resource Allocation
– Access Points are aware of the traffic demand.
– 802.11e compliant AP’s and nodes are necessary for
QoS differentation.
– AIFS, CWMin are among the parameters that can be
controlled.
– We use AIFS as the control parameter for our
simulations in ns-2.
– The end user application is VoIP.
• Modified VoIP Utility Function
– MOS*(R) = MOS(R) - ( - 1)
64. Resource Allocation Algorithm for WLAN
• For our example, we consider social distances to be chosen from the
set {1,2,3} with 1 signifying the highest priority.
• Control parameter = AIFS
CHOOSE AIFS
Max Unew(AIFS, ) – cost(AIFS, )
Subject to
1<=AIFS<=7
1<= <= 3
Algorithm computes Loss (L)
loss and delay for the And
current mix of calls after Delay (D)
matrices
adding this new call
or
67. Real World Implementation
• Effective Distance
– To measure this quantity, applications need to
become location-aware.
– Social distance awareness is also necessary. But this
is usually easier, since it is determined by the user
herself.
• Our Solution
– Implement a Wi-Fi Positioning System for locating
devices when inside buildings (EB-II).
– Devices are GPS-enabled (iPhones/Android devices)
to facilitate positioning when outside.
68. Wi-Fi Positioning System
• Steps of positioning
system
1. Client retrieves data
(Visible Access
Points and their
RSSI)
• 2.Client sends data to
server
• 3-6.Server enters data into
database, uses algorithm to
calculate position
• 7. Other Clients open
map using browser and get
the location information
from the server
69. The V911 Application – A Location
Aware Application
• Emergency response application, with the
locations being determined using WPS.
WPS
Server
User Helper
70. Ongoing Work
• Implement applications which have a social
context in addition to being location-aware.
– A game with 4 teams competing against each other.
• Perform trials with devices spread out both
indoors and outdoors.
• Measure network performance metrics (delay,
loss, jitter etc).
• Relate the user’s experience to the metrics –
culminating in the definition of the social distance
aware utility function for this application.
71. Traffic Modeling
Motivation:
Better models required for performance studies in:
QoS, admission control, testing
Goals:
Accurate models of statistical behavior of traffic
Computationally efficient models of traffic
72. Purpose of Traffic Measurements
Capture traffic from different places across
research and education campuses.
Identify emerging traffic patterns.
Deconstruct and analyze the intersecting
networking and social distances (K-20).
Calculate end-to-end delay and packet loss
rate based on traffic measurements.
73. Design of the Broadband NGN Architecture
Quadruple Play Service Delivery Architecture
= ?
Triple Play Services (Voice, Video, Internet) + Broadband Wireless Access
Services
Single-Edge Multi-Edge
Geographic
Separate devices for various
All services flow through single device, services. Could be service specific
Centralized
located in a centralized PoP edge, or common per-subscriber
PEP but on multiple systems
All services flow through a single device, Some services are produced on
Distributed distributed in the architecture close to the distributed devices, whereas other
subscriber services are produced centrally
Clustered: Multiple Edge routers will support one service
Un-clustered: One Edge router for one service
74. Network Design Problems
• Horizontal Design
– Location Problems: “Pure” topological design
problems where demands are not taken into account.
More of where to place the devices
– Dimensioning Problems: Network Design Problems
that take into account the demand volumes. More on
dimensioning the locations and interconnections
• (Proposed) Vertical Design
– Optimally Process next generation services/flows
– The network devices may support multiple layers (IP
routing, SON)
75. Centralized Single Edge vs Multi-Edge
Architecture
• Single Edge
– All types of traffic are
backhauld to the BNGs
located at a single PoP
– Subscriber termination
functionality, and IP
Qos policies are
executed in the BNG
– Multicast Traffic for
video is transmitted
from the edge router
(L2 multicast VLAN)
• Multi Edge
– Different types of edge
routers (BNG/Video
BNG and business at
MSE)
– Clustered: Multiple
number of BNGs and
MSEs at the same PoP
– Benefits from
incumbency (easier to
evolve)
77. How to measure traffic?
Collection of traffic statistics is currently performed
• Flow monitor, e.g., Cisco NetFlow
• Sophisticated Network Monitoring Equipment
The port based measuring system (identify applications
based on ports) captures only 60% of traffic
Our implementation for netflow data:
78. Network monitoring through Netflow
data
• Proof of concept for netflow data: Classification Apps are on average ~60%
• http://benediction.ece.ncsu.edu/ece480/
79. Traffic hiding into similar ports
• Applications do not belong to static ports
• Maybe they belong to P2P applications that use dynamic
ports (even FTP does initiation at 21 and jumps to 20;
Skype also uses 80)
• Maybe they belong to media applications
It is hard to find out who generated what unless expensive stateful (resource
draining) DPI monitors are used
• The proposed implementation
• Level 1: Signature checking in Packet files (processor intensive)
and store in hash table
• Level 2: Check every flow in the hash table
• Worst case: Check for Port and ToS on the header
81. Our Vision for “Weather Maps”
• Share statistics among
NCREN research community
• Anonymize data
• Multiple measuring points
• Per link traffic/Per application
(taking a step forward from
regular MRTG/CACTI as
shown in the figure)
• http://benediction.ece.ncsu.edu/
82. Methodology for NCREN Related Campuses
• Detailed per-application capture in various
locations of NCSU campus.
• Analysis of the intersecting social, networking
and geographical distances of K-20 users.
• Identify architectural modifications to improve the
overall performance of the network.
• Propose how new wireless access technologies
and mobility patterns affect the traffic.
• Analyze flow of IP packets through an
aggregation network to calculate probability of
end-to-end delay and packet loss.
83. Peer2peer Networks (e.g., eDonkey,
emule, Gnutella, Kazaa)
• Peer-to-peer file sharing applications
have evolved to one of the major traffic
sources in the Internet.
• In particular, the eDonkey file sharing
system and its derivatives are causing
high amounts of traffic volume in today's
networks.
• The eDonkey system is typically used
for exchanging very large files like
audio/video CDs or even DVD images.
• Peer2Peer Network based on
“torrents”: function different than other
P2P programs. A large file is divided into
chunks. Peers interested in the same file
self-organize into a torrent and peers
exchange file chunks with each other.
84. Modeling Game Traffic
• In order for a game with client-server topology to be effective it must
accommodate huge lag (ping, roundtrip delay) and loss
• Methods to solve the lag in games:
1. Client-side prediction of the game state i.e. movements of objects
and other players
2. Combine movement with inertia or reducing maximum velocity of
objects prediction is even more effective
The most robust games tolerate lag up to 1 sec (mean value is usually
around 200ms) and loss up to 40%
Quality based on Ping times
<50ms Excellent
<100ms Good
>150ms Bad
Generally is based on the robustness and type of the game (there are
games that accommodate lag of 200ms)
85. Modeling Games an Open Issue
• Games have different requirements (a 3D first
shooter game requires very low, but 2D strategy
games can accommodate higher lag)
• Clients do not act independently (social
networking aspects)
• Server traffic per client is dependent on all clients
(many models assume independency)
• Each game sends different messages and there
is not currently a general social formalization.
• Will 3D Virtual Worlds become the “Web 3.0”?
86. Optimization Model
The following linear integer programming problem needs to be solved:
min
n{i , j , k } c x
co x ,cYn x ,c co L 3Y L 3n ,t co L 2Yn L 2
t
c [1G,10G ], x [ L 2, L3]
• Yn x ,c Number of interfaces with capacity c at location n
• Yn x Number of elements (x=L2 switch, x=L3 router)
Constraints
• Quantitave (Elements/Interfaces)
• Qualitative (placement of functionalities)
• Number of Ports/Interfaces
• Line Card Capacity per Network Element
• Interface Capacity (1Gbps/10Gbps)
• Traffic Flow constraints (local or non-local)
87. Capital Expenditures (CapEx)
• Argument around number and cost of devices
• Pure L2 aggregation networks combined with
centralized edge architectures seem to have cost
advantages
– BUT: how to handle multicast traffic in a pure L2 aggregation
network?
– Motivates IP-enabled aggregation networks
• Number of IP enabled devices in centralized and
distributed architectures of same order of magnitude
– Thus Capex in a similar range
– CapEx of distributed architecture comparable to centralized
architecture with IP-enabled aggregation infrastructure
88. Summary: Service Oriented Networks
• Our work focuses on:
– Capturing “presence” and location awareness
– Quantifying socio-technical interactions
– Characterizing workload spatially and temporally
– Optimizing over multiple resources, across
layers
SON and
Convergent Nets
“Cloud”
Social Apps and Workload Aggregation
Virtual and Distributed
Collaboration Delivery