Transcript College of Engineering and Science

Intelligent Systems (IS)
Computer Systems Architecture (CSA)
Focus Areas
Introduction for Prospective Graduate Students
Ian Walker
Fall 2012
Outline
 Who and what are we?
 Classes, requirements, planning
 Funding opportunities, assistantships
 Degree options
 Sample research projects
 Q&A
Who are we?
 Loose confederation based upon common research interests
Loose mission statements:
IS: Building smarter machine systems
CSA: Building better/faster computing machines
Who:
IS (9 Professors): Birchfield, Brooks, Burg, Dawson, Groff, Hoover,
Schalkoff, Venayagamoorthy (new!), Walker
CSA (9 Professors): Birchfield, Brooks, Gowdy, Hoover, Ligon,
Schalkoff, Shen, Smith, Walker
Who are we?
Current enrollment
IS: 30-50 graduate students
CSA: 10-25 graduate students
Lab space
IS: Riggs 10, Riggs 13/15/17 (main lab), EIB 258 (main lab)
CSA: Riggs 309 (main lab), EIB 352 (main lab), Cluster room
Shared: Riggs 315/7, EIB 341, ...
Sample Research Areas
 Sensor networks
 Tracking filters and embedded systems
 Physiological monitoring systems
 Nonlinear system modeling and control
 Audio and visual spatial sensing
 Biologically inspired robotics
 (More that are not listed here)
Classes (IS)
Required (all these courses offered once per year) :
ECE 801 - Analysis of Linear Systems
ECE 847 - Digital Image Processing
A 600-level course chosen from (642, 655*, 668)
One of (854, 855, 856, 868, 869, 872, 874*, 877)
*For
Computer Engineering, 649 replaces 655, and 874 is removed from list
Other IS courses (typically offered once per 3 semesters):
804, 805, 854, 856, 872, 893 (various)
courses from other focus areas or departments are allowed
Planning: Take core early, figure out what you would like to do
See p. 35 of http://www.clemson.edu/ces/departments/ece/document_resource/grad/Grad_Student_Handbook_2011
Classes (CSA)
Required:
A software course (ECE 617, 852, 855, or 873)
An architecture course (ECE 629, 668, 842, or 851)
A networks course (ECE 640, 649, 848, or 849)
Other CSA courses:
any from the above lists
courses from other focus areas or departments are allowed
Note: 693 and 893 are used for new courses. Be sure to sign
up for the right section number.
See p. 32 of http://www.clemson.edu/ces/departments/ece/document_resource/grad/Grad_Student_Handbook_2011
Advisors
Selecting a faculty advisor is a two-way decision
All faculty use different criteria for evaluating students
Performance in core course taught by that professor
Evaluation of volunteer or startup work in lab
Probationary period
Assistance to PhD or senior graduate student
Funding
Grading Assistantship (GA) - assist prof. with a course
Teaching Assistantship (TA) - teach lab sections
Research Assistantship (RA) - assist prof. in funded project
GAs and TAs are administered by department
RAs are generally offered to PhD students, or sometimes
masters students showing potential and commitment for
PhD
You do not need funding to get involved in research
Degree options
Majors (at masters and PhD level):
Computer Engineering (CpE)
Electrical Engineering (EE)
Options:
Focus area (IS is one of six areas in department)
Non-thesis (coursework only)
33 hours (11 courses)
Thesis
30 hours (8 courses + research)
best to examine options after first semester completed
typically work with PhD student
probably adds a semester - 2 years total
Direct-PhD
60 hours (14 courses + research)
saves 2 courses compared with Masters + PhD
possible to get an MS along the way
For details, see http://www.clemson.edu/ces/departments/ece/document_resource/grad/Grad_Student_Handbook_2011
Recent graduates
Ph.D. students - Academic Positions
Clarkson University at Potsdam, New York
University of Michigan at Ann Arbor, Michigan
Louisiana State University, Louisiana
University of Florida
Ph.D. students - Industrial Positions
Lucent Technology in Connecticut
Oakridge National Laboratories in Tennessee
Mayo Clinic in Minnesota
MS Students - Ph.D. Pursuits
German Aerospace Institute in Germany
Stanford University in California
MS Students - Industrial Positions
General Electric in Virginia
IBM in North Carolina
Intel in Columbia and San Francisco
Yahoo! in California
Harris in Florida
GM-Fanuc in Michigan
Name: Kumar Venayagamoorthy
Focus Area: Power/IS
http://www.people.clemson.edu/~gvenaya/
Research Area:
Real-time power systems
Current Projects:
Smart Grid research
Name: Richard Brooks
Focus Area: CSA/IS
http://www.clemson.edu/~rrb
Research Area:
Distributed Systems / Information Assurance / Coordination
Current Projects:
AFOSR – Detection of Tunnelled Communications Protocols
Industry – Data Leak Prevention
NSF – Network Security Experimentation with GENI
Department of State – Internet Liberty Support for West Africa
Relevant courses:
ECE449 / 649
Data Leak Prevention (DLP) solutions monitor and control data flow
Current DLP solutions are syntax based
- Hash functions
- Regular expressions
- Keyword search
We focus on data semantics
Singular value-based approach
Apply singular value decomposition to term-document matrix.
Find concepts by retaining a number of dimensions.
Hidden Markov Model (HMM)-based approach
Build HMMs based on terms we retained in singular value-based method.
Singular value
decomposition
Find transition probabilities of each document and estimate the probabilities of unobserved transitions.
Probabilistic Context-Free Grammar (PCFG)-based approach
Obtain parse trees of sentences in training documents.
Identify features in the parse trees.
Transmission
Cache
VLSI
…….
Distributed Denial of Service Attack (DDoS) Analysis
WiMAX BCR System Parameters and DDoS Attack Analysis
•
Factorial Experimental Design and ANOVA analysis of avg. throughput Ns-2 simulator used for software
simulations
Real software-defined radio testbeds used for hardware simulations
1%
8%
4%
NS-2 Software Simulations
7%
0%
0% frame_duration,
0% X1
22%
31%
3%
6%
number_of_attac
kers, X2
dos_backoff_star
t, X3
18%
•
1%
0% 4%
1%
Setup the network using Clemson University GENI resources.
Use Operational Network traffic.
Generate DDoS attack traffic using Clemson Condor Cluster.
Analyze performance of DDoS detection methods.
11%
Backoff_start, X1
56%
Throughput of Indoor
Performance Analysis of DDoS Detection Methods on Operational Network
-
42%
Backoff_end, X2
X1*X2
Other
85%
Throughput of Outdoor
•
•
•
A bootable USB drive with the Linux system will access the proxy network.
The proxy network deploys botnet which changes DNS and IP address to avoid detection and tracking.
With this, the democracy advocates, NGOs, and journalists are protected from network censorship and
surveillance.
Detecting Hidden Communications Protocols
•
Protocol analysis of Tor through side-channel attacks
–
–
–
•
Protocol represented as a hidden Markov model (HMM)
Side-channel information: delays between packets
Using zero-knowledge HMM inference algorithm to rebuild the model, i.e. the protocol used by A.
Botnet traffic detection
-
Infer HMMs from botnet timing data
Use confidence interval approach to detect botnet traffic
Result: 95% TP and 2% FP
Name: Melissa Smith
Focus Area: CSA
http://www.parl.clemson.edu/~smithmc/
Research Area:
High-Performance Reconfigurable Computing/ Heterogeneous Computing
Current Projects:
Heterogeneous Mapping and Acceleration of Scientific Algorithms
Acceleration of Gene Co-Expression Network Generation
Performance Models for Hybrid Computing
Exploration of Concurrent Biometric Algorithms for Emerging Reconfigurable Architectures
Relevant courses: (ECE 668, 845, 842, 873, 893)
SNNs Optimizations with Multi-Core
Architectures
Spiking Neural Networks (SNN): preferred
neural network models for simulating the
biological behavior of a neuron
Ultimate goal of scientists:
Izhikevich’s Model
Flop/Byte : 0.65
Wilson model: Flop/Byte: 0.86
Level 1
Morris Lecar Model Flop/Byte:4.71
HH Model Flop/Byte : 6.02
HH model Speedup for different Architectures
900
Fermi GPU, OpenCL
800
Speedup
Model mammalian brain activity
(1011 neurons – 1014 synapses)
Object recognition/identification
Two-level character recognition network
Level 2
w/ two SNN models:
Fermi GPU, CUDA
700
Telsa 870, OpenCL
600
Telsa 870, CUDA
500
AMD GPU, OpenCL
Intel Xeon
400
AMD Opteron
300
IMB PS3
200
100
0
0
2
4
6
Neurons (millions)
Results published in HiCOMB’10, Journal of Supercomputing, & Concurrency and Computation
Exploring Multiple Levels of Heterogeneous
Performance Modeling
Synchronous Iterative GPGPU Execution
(SIGE) model
Regression models for
Use Synchronous Iterative GPGPU Execution
CPU/GPU computations
using Algorithm FLOPS (SIGE) Model for Synchronous Iterative
Algorithms (SIAs)
and Bytes
Relevant Equations describing the SIGE
Model
Texecution = ∑Tcomp. + ∑Tcomm.
Tcomp.= Tpre-process + Tpost-process + TCPU + TGPU
TGPU = TGPU-Kernel + TPCIE-Transfers
TPCIE-Transfers = Thost-to-device + Tdevice-to-host
Tcomm. = ∑Tnetwork-transactions
Initial validation of low-level abstraction
model for GPGPU clusters
Regression-based performance prediction
framework
SIA case studies: Spiking Neural Network
Regression models for PCIE
and Infiniband using micro- (SNN) models
benchmarks
Achieved over 90% prediction accuracy
Gene Co-Expression Network
Construction
Correlation Matrix
Calculation
Threshold Calculation
40
Running Time (Hours)
Running Time (Hours)
25
20
18X Faster
15
10
5
0
C
OPT
10
Accelerating construction of
gene co-expression
networks, which analyze the
relationships among
thousands of genes
Previous techniques were
slow and use excessive disk
space
JAVA
Data Storage Requirement
40
•
35
7X Smaller
30
File Size (GB)
•
20
0
JAVA
•
45X Faster
30
25
20
15
10
•
5
0
JAVA
C
SYMM
BIN
C
OPT
Our acceleration has allowed
generation of hundreds of
gene networks of multiple
sizes and types (rice, yeast,
and human) for in-depth
analysis never before
possible
Future work with GPUs and
other accelerators will
provide additional
Robust Facial Recognition with Highly-Parallel Architectures
The rapidly growing field
of biometrics uses
physical features to
perform identity
authentication.
Several facial recognition
algorithms have been
developed that can adapt to
particular types of image
variation, but no single
algorithm can provide robust
identification.
Facial recognition is the
user’s most convenient
biometric but often suffers
from poor performance,
especially in applications
with wide image variation.
FPGAs and GPUs provide the
necessary parallelism to run
multiple algorithms
simultaneously and fuse their
results together to enable
accurate recognition.
Name: Walt Ligon
Focus Area: CSA
http://www.parl.clemson.edu/~walt
Research Area:
Parallel Computing, Parallel File Systems, Programming Environments
Current Projects:
Parallel Virtual File System (PVFS)
High End Computing I/O Simulator (HECIOS)
Relevant courses: ECE 851, 873, 329, 493 (MPI)
Name: Robert Schalkoff
Focus Area: CSA/IS
http://www.ece.clemson.edu/iaal/index.html
Research Area:
Soft Computing/Parallel Programming
Current Projects:
An algebraic framework for multi-class motion estimation
using unsupervised learning with GPU implementation
Relevant courses: ECE 856, ECE 855, ECE872, ECE 642, ECE 847
An algebraic framework for multi-class motion estimation
using unsupervised learning with GPU implementation
Optical flow constraint equation (OFCE) is
Ix* u + Iy* v + It = 0
Pixel locations that suffer aperture
problem have rank-deficient system.
The min-norm solution of rank-deficient
system leads to motion estimates with low
confidence. High confidence is associated
with vectors that do not suffer aperture
problem.
Motion vectors (u,v) are separated into
two sets; one set of vectors (Hp) that
suffer aperture problem and another set
of vectors (Hc) that do not.
Implementation with NVIDIA CUDA: Compute Unified Device Architecture
SM 0
Kernels for Motion Estimation:
1. Gradients
2. Local Motion
3. SOFM/NG
SP SP
SP SP
Shared
Memory
SP
SP
SP
SP
1
Texture
Memory
Global
Memory
SM 0
SP SP
SP
SP SP
SP
Shared
Memory
2
A
Mutable
B
C
B
A
C D
Immutable
SP
SP
D
Name: Haiying Shen
Focus Area: CSA
http://www.ces.clemson.edu/~shenh
Research Area:
Distributed computer systems and computer networks
Current Projects:
Leveraging Hierarchical DHTs and Social Networks for P2P Live Streaming
P2P File Storage and Sharing System for High-End Computing
Pervasive Data Sharing Over Heterogeneous Networks
File Replication and Consistency Maintenance in Pervasive Distributed Computing
Hybrid Wireless Networks
Self-organizing P2P-based File Storage System in HPC
Relevant courses: ECE 429/629, ECE 893
P2P Live Streaming/VoD
Social network
• Internet-based video streaming
applications attract millions of
online viewers every day.
C
A
• The incredible growth of viewers
and dynamics of participants have
posed a high quality-of-service
(QoS) requirement.
channel B
cluster
• Goals: high scalability, availability,
low-latency.
DHTs
(Channels)
channel
cluster
n
channel
cluster
Pic from http://www.fmsasg.com/SocialNetworkAnalysis/
Features:
(1) Distributed Hash Table is constructed for
content delivery to increase scalability,
availability
(2) Social network is used for accurate conten
recommendation and channel switch to reduc
video delivery latency
(Images captured from paper Flexible Divide-and-Conquer protocol for multi-view peer-to-peer live streaming,
Preliminary results published in ICPP10, Infocom11, IEEE TPDS 11
P2P’09)
GENI Experiments on P2P, MANET, WSN
Networks
We will implement three existing data sharing algorithms on the P2P, MANET and WSN networks, thus
identify and investigate potential issues in the data sharing applications in heterogeneous networks.
Data sharing in P2P networks
(Cycloid P2P)
Locality-based distributed data sharing
protocol (LORD) in MANETs
Features:
(1) Energy-efficient & scalable.
(2) Reliable & dynamism-resilient.
(3) Similarity search capability
Features:
(1) Constant maintenance overhead regardless of
the system scale.
(2) Scalability, reliability, dynamism-resilience, selforganizing.
Number of
nodes:
100
Dimension:
6
Node failure
rate:
0.1-1 natural
Lookup/Insert
interval
10-100s to
every node
Number of nodes (ORBIT)
100
Total lookups
10000
Moving speed dist. (m/s)
[0.5-2.5], [1-5], [20-30]
Spatial-temporal similarity
data sharing (SDS) in WSNs
Number of sensors
128
Node in zone
9
LSH destinations
5
Features:
(1) Efficient spatial/temporal similarity data storage.
(2) Fast query speed.
(3) Low energy consumption.
Leveraging P2P in HPC/Cloud Computing

P2P network is well-known for scalability,
reliability and self-organizing





P2P-based Resource Management


Effective and efficient P2P content delivery
algorithm design (TC11, TPDS10,
INFOCOM11, IPDPS08)
P2P-based Reputation Management




Social network based P2P overlay
construction (under review of INFOCOM12 )
Locality aware P2P overlay construction
(CCNC 09)
Interest aware P2P overlay construction
(CCGRID 09)
User behavior pattern aware P2P overlay
construction (In preparation for IPDPS 12)
Social network Collusion detection (IPDPS11)
Spam filtering (INFOCOM11)
Game theory based cooperation incentive
analysis (ICCCN09, TMC)
Grid computing
P2P-based File Storage System in HPC


File replication (JPDC09)
File consistency maintenance (TPDS11)
Pic from http://innovationsimple.com/web-hosting/cloud-hosting-web-hosting/benefits-of-cloud-computing/
Cloud computing
Name: Darren Dawson
Focus Area: IS
http://www.ece.clemson.edu/crb/welcome.htm
Research Area:
Nonlinear Control and Estimation for Mechatronic Systems
Current Projects:
Following 3 Slides
Relevant courses: ECE 874, 801
Visual Servoing of Robot Manipulators
 Problem: Control of Moving Objects in an Unstructured Environment is Difficult
due to the Corrupting Influences of Camera Calibration with regard to Task
Planning
 Solution: Close the Control Loop with
Camera Measurements
 Testbed Features a High-Speed
Real-Time Camera System
 2.5D Visual Servoing
 Design a Controller to Regulate
the Position and Orientation
of the End-Effector
 Control Strategy Uses Both 2D
Image-Space and 3D Task-Space
Information
Next Generation Hardware-in-the-Loop
Ground Vehicle Steering Simulator
Custom Honda CRV steering
simulator with electric servo-motors
Test platform supports development of
advanced ground vehicle steering
technology using concepts from
“robotics” field
Also examining in-vehicle operator
feedback channels
• Visual (scene, lights)
• Haptic (steering wheel, …)
• Audio (tones/chimes/voice)
Human subject testing
Advanced Automotive Thermal Management Systems - Smart
Components
 Goal is to improve the engine’s cooling/heating
system operation using mechatronic technology
•
Improved fuel economy
•
Reduced tailpipe emissions
•
Flexible thermal system design
•
Enhanced control of engine temperatures
 Replace mechanical cooling system equipment
with electric/hydraulic-driven components
 Develop mathematical thermal models
Name: Tim Burg
Focus Area: IS
http://www.clemson.edu/~tburg
Research Area:
Nonlinear Control Applications
Current Projects:
Unmanned Aerial Vehicles
Biofabrication
Haptics
Environmental Monitoring
Relevant courses: ECE 874, 801
Bioprinting
 Bioprinting - an approach to tissue engineering
 Cells are precisely placed in a 3D structure using
inkjet printer technology.
 Active collaboration with Bioengineering.
 ECE research focused on system integration,
modeling, and control.
Haptics
 Objective Is to identify, demonstrate, and quantify the
potential benefits of specialized haptic user interfaces
within a collaborative environment.
Name: Stan Birchfield
Focus Area: IS
http://www.ces.clemson.edu/~stb
Research Area:
Computer Vision
Current Projects:
Vision-based mobile robot navigation
Vehicle traffic monitoring
Robotic laundry handling
Relevant courses: ECE 847, 877, 904
Vision-Based Mobile Robot Navigation
 Mobile robot equipped with single, off-the-shelf inexpensive camera
 Developing algorithms for
 Traversing a known path by comparing the coordinates of tracked feature points
 Detecting doors in indoor environments for navigation
 Following a person moving about the environment, maintaining a desired distance
 Applications: courier robots, tour guides, physician assistance
Vehicle Traffic Monitoring Using Cameras
 Developing algorithms for detecting,
tracking, and classifying vehicles
automatically using video
 Low-angle cameras cause occlusion and
spillover
 Shadows, reflections, and environmental
conditions are addressed using a
combination of feature tracking and
pattern detection
 Applications:
 intelligent transportation systems (ITS)
 incident detection and emergency
response
 data collection for transportation
engineering applications
Adam Hoover
Focus Area: IS/CSA
http://www.ces.clemson.edu/~ahoover/
Research Area:
Tracking systems, embedded systems
Current projects:
See the next 2 slides
Relevant courses: ECE 854, 668
Bite Counter
1 in 3 Americans is obese, another 1 in 3
is overweight; worldwide there are more
overweight than underfed people
Worn like a watch
Automatically tracks how many
bites of food have been taken
•
•
•
Bite count vs calories for 54 meals
2011-2012 large cafeteria experiment in main campus dining hall
Equipment and software for recording and correlating video, scale, gyroscope data
Signal analysis to improve bite detection accuracy and bite:calorie correlation
Ultrawideband Position Tracking
same
idea
Trilateration measures distances from a set of transmitters to a receiver to calculate position.
•
•
•
Ubisense system in Riggs basement
Particle filter methods to improve accuracy
Noise modeling, combination with other sensors
and other sources of information such as maps
Richard Groff
Focus Area: IS
http://www.ces.clemson.edu/~regroff
Research:
Robotics and control applications at small length scales
Computational and Experimental Tissue Modeling
Biomimetics
Current Projects:
Synthetic butterfly proboscises
Biofabrication and Tissue Modeling
(under revision)
Relevant coursework:
ECE801 (linear systems), ECE847 (digital signal processing)
for some projects, some background in magnetostatics, solid mechanics,
materials science, and/or molecular biology desired
Synthetic Butterfly Proboscis
Proboscis
Experimental
Platform for
Magnetic
Microfibers
 Butterflies can drink fluids of widely varying
viscosities by controlling the shape of their
feeding tube (probosicis)
 Using custom fibers from Materials Science
Department, generate a synthetic proboscis that
can sample widely varying fluids
 Fibers are paramagnetic or piezoelectric
 Control fiber shape using magnetic or electric
fields
 Preliminary work on modeling and position control
of magnetic microfibers
Tissue Engineering via Biofabrication
Fluorescent-dyed murine D1 mesenchymal stem cells (red)
and murine mammary cancer cells (red)
 Biofabrication – develop a system to place living
cells in 3D patterns mimicking native tissue
 many subprojects
 Develop computational model for interaction of
tumor cells and epithelial stem cells
 “Tissue Description Language”
 Specify Describe initial condition for
computational model
 Specify structure for biofabrication
 Use TDL to study systems biology problems in
cancer. (Feedback via intercellular signalling)
Name: Ian Walker
Focus Area: IS/CSA
http://www.ces.clemson.edu/~ianw/
Research Area:
Robotics
Current Projects:
Trunk and tentacle robots
Intelligent Robotic Workstations
Relevant courses: ECE 655, 868, 869
Invertebrate’ robot trunks/tentacles
Animated Architecture
 Integrate Robotics and Architecture
 Goal “Animated Work Environment”
What should you do next?
Find out more about specific research projects
web, senior graduate students, faculty
Contact potential advisors about projects, openings
faculty attending this meeting may be recruiting currently
Either
a) Mutually agree on advising relationship
OR
b) Establish criteria for being evaluated/considered
OR
c) Seek another advisor/project
Q&A
?