F3: Application Case Studies

1
F3 Application Case Studies HLLs

• Dr. Alan D. George
• Professor of ECE, University of Florida
• Dr. K. Clint Slatton
• Asst. Professor of ECE CCE, University of
  Florida
• Brian Holland
• Ph.D. Student, University of Florida

2
Outline

• Project Goals, Motivations, Challenges
• Background and Related Research
• Project Team Members (faculty students)
• Y1 Tasks
• Overview of Y1 Tasks
• Task 1 Select set of challenging diverse
  mission scenarios
• Task 2 Design, develop, analyze, refine, and
  evaluate multi-level parallel algorithms for
  selected mission scenarios
• Task 3 Explore design options tradeoffs with
  HDLs and HLLs
• Task 4 Formalize insight from tasks to theorize
  broader implications
• Y1 Milestones, Deliverables, Budget
• Conclusions, Member Benefits

3
Project Goals, Motivations, Challenges

• Goals
• Develop understanding from case-study experience
  of decomposition mapping strategies w/ complex
  apps
• Scenario applications defined jointly with
  members
• Hardware/software partitioning, co-design,
  optimization
• Concomitantly explore complimentary issues
  including fault tolerance (e.g. ABFT v. NMR),
  design portability, and precision
• Motivation
• High-performance reconfigurable computing is
  still in its infancy
• Field needs more knowledge, insight, proof of
  success w/ real apps
• Research Challenges
• Multilevel algorithm partitioning, analysis, and
  optimization
• Design tradeoffs for HLL vs. HDL for application
  development
• Balancing performance with portability,
  precision, dependability

Wheres the beef?
4
Background Related Research

• F3 The Application Salad Bar
• 3 components per mission scenario
• Application
• System platform
• Parameters
• What follows is a sampling of applications and
  parameters recently under study at UF
• Adaptive signal processing applications
• Space-based radar processing applications
• Algorithm-Based Fault Tolerance (ABFT)
• High-Level Language (HLL) paradigms
• HPC cases, HPEC cases, dual-purpose

Mission scenarios might range from neutron
science on HPC for ORNL to space science on HPEC
satellite for NASA
This is a sampling, but by no means an
exhaustive list of potential applications to
explore.
5
Background Related Research

• Embedded Remote Sensing Apps
• On-board processing of lidar (light detection and
  ranging) data would allow near real-time
  environment assessment and in-flight adjustments
• UF owns airborne and vehicle-based lidars
• Building extraction, ground point calculation,
  DEM (Digital Elevation Model) generation,
  vegetation filtering and data compression
• Other applications change/target detection,
  scene analysis, image/video segmentation, clutter
  rejection

DEMs of forest canopy tops and filtered lidar
penetration to ground
6
Background Related Research

• Non-Embedded Remote Sensing
• Non-parametric PDF estimation is critical to
  statistical decision theory and probabilistic
  classification
• Topography estimation (through foliage)
  Hierarchical data segmentation and Bayesian
  networks
• Soil drainage classification high-dimensional
  terrain classification driven by information
  theory and boosting algorithms
• Multi-resolution, multi-sensor data fusion for
  improved classification and replacing of
  corrupted or missing data

7
Background Related Research

• Space-based Radar Applications
• Ground-Moving Target Indicator (e.g. 512 MB
  typical data set)
• Tracks moving targets on ground from air or space
• Continuous, real-time processing
• Synthetic Aperture Radar (e.g. 2 GB typical data
  set)
• High-resolution images of ground from air or
  space
• Less-stringent processing deadline (relative to
  GMTI)
• Hyperspectral Imaging (e.g. 4 GB typical data
  set)
• Target detection and classification in
  hyperspectral images
• Incredibly large data sizes, relatively simpler
  computations

Hyperspectral Imaging
Ground-Moving Target Indicator
8
Background Related Research

• Fault Tolerance with RC
• Classical approach to mitigate SEU (single-event
  upsets)
• TMR Triple-Modular Redundancy provides nearly
  perfect protection against SEU but at hefty cost
  of area and power consumption
• Periodic configuration scrubbing to repair errors
• Leverage other research, explore RC advantages
• Partial or Selective TMR
• Autonomous repair cells
• Classical approach to mitigate computational
  errors at algorithm level
• ABFT Algorithm-Based Fault Tolerance is an
  approach providing ability to detect and fix
  errors by incorporating redundant data into the
  algorithm

Leverages on-going research at Florida with
Honeywell in fault-tolerant computing for space.
9
Background Related Research

• High-Level Languages

• HLLs vs. HDLS
• High-level languages have potential to provide a
  rapid development path for FPGAs
• Various languages leveraging C, Fortran, and
  Matlab can be used (depending upon compiler)
• Much research has been underway but more is
  necessary with mainstream applications

• Challenges
• Which language and compiler are best for platform
  and application of a given scenario?
• Using right tool can be very beneficial using
  wrong tool can be far worse than VHDL
• Using a high-level paradigm does not replace
  human interaction or design
• Each language has its own nuances

For example, MAPLD05 paper by Holland et al.
from U. Florida.
10
Project Team Members

• Faculty
• Dr. Alan George, Professor of ECE
• Dr. Clint Slatton, Assistant Professor of ECE and
  CCE
• Dr. Herman Lam, Associate Professor of ECE
• Dr. Dapeng Wu, Assistant Professor of ECE
• Students
• Brian Holland student project leader
• 2nd-year doctoral student, University of Florida
• BS in ECE, Clemson University, 2005
• Alumni Fellow
• Daniel Espinosa
• 2nd-year masters student, University of Florida
• BS in ECE, University of Florida, 2005
• Internship at NASA Goddard Space Flight Center
• Karthik Nagarajan
• 2nd-year masters student, University of Florida
• BS in ECE, University of Madras (India), 2003
• TBD
• 0-3 additional graduate students

11
Overview of Y1 Tasks

• Task 1
• Select set of challenging scenarios
  (apps-platforms-parms) with maximal intersection
  of (a) member interests, (b) faculty expertise,
  and (c) suitability for success with RC
• HPC scenarios, HPEC scenarios, or cases
  reflective of both
• Task 2
• Design, develop, analyze, refine, and evaluate
  multi-level parallel algorithms for applications
  in selected scenarios
• Note this large task will be divided into
  subtasks, one per scenario or app
• Task 3
• Explore options and tradeoffs with HDLs, HLLs,
  and mixes for each app and its modules in
  implementation of new parallel algorithms on
  target platforms
• Task4
• Formalize insight from previous tasks to theorize
  broader implications for other suites of problems
  as well as a taxonomy of application classes

12
Task 1

• Select a set of challenging and diverse mission
  scenarios
• What scenarios and applications are of keen
  interest?
• Primary focus will be apps of prime interest to
  our members
• Some of many app areas where UF faculty have
  expertise
• STAP, SAR, GMTI, and other SBR and related apps
• Multi-scale image fusing (e.g. multiscale Kalman
  smoothing)
• Multi-rate signal processing (e.g. airborne laser
  swath mapping)
• Image processing (e.g. 3D reconstruction from
  LIDAR images)
• Platforms to target in mission scenarios?
• Broad range available in our lab, for HPC or for
  HPEC
• What parameters should be explored?
• Total emphasis on performance, speedup, and
  scalability?
• Balancing performance with resource and power
  limitations of devices?
• Sacrificing some performance for portability and
  adaptability?
• Incorporating unique challenges such as fault
  tolerance?

13
Task 2

• Design, develop, analyze, refine, and evaluate
• multi-level parallel algorithms
• Performance
• Investigate dual-paradigm architectures and
  algorithms that provide optimal performance for
  each mission scenario
• Explore tradeoff issues in hardware/software
  partitioning, co-design, and optimization, as
  well as numerical precision amass share tools
  insight
• Fault Tolerance
• Investigate tradeoffs between NMR and ABFT
• Determine classes of RC applications most
  amenable to ABFT
• Standardization and Portability
• Performance is often tightly linked with
  mission-specific applications customized for
  target platform architecture
• Adaptation to new systems is generally performed
  manually and can require significant time and
  application modification
• Balancing performance with portability and
  adaptability will be necessary for some mission
  scenarios

One of our papers on this subject was awarded
as Distinguished Research Paper at ERSA06.
14
Tasks 3 4

• Task 3 Explore tradeoffs with HDLs and HLLs
• Explore tradeoffs between traditional HDLs and
  new HLL approaches for performance and efficiency
• Evaluate via case-study experience then determine
  and best leverage unique benefits of available
  languages tools
• e.g. VHDL, DIME-C, Impulse C, Handel C, AccelDSP,
  Mitrion, et al.
• Task 4 Explore broader implications of case
  studies
• Initial mission scenarios will be focused upon
  specific applications, selected platforms, and
  target parameters
• However, concepts learned are not necessarily
  unique to each case broader conclusions can be
  explored and drawn
• Concepts such as standardization, adaptability,
  and portability may also play an important role
  in an applications broader implications

15
Y1 Milestones, Deliverables, Budget

• Milestones
• 02/01/07 Finish scenario selection definition
  (Task 1)
• 11/01/07 Finish case studies and tradeoffs
  (Tasks 2-3)
• 12/01/07 Finish study of broader implications
  (Task 4)
• Deliverables
• Midterm and final reports documenting research
  methods, progress, results, and analysis
• All developed core and application codes
• One or two scholarly conference and/or journal
  publications
• Budget
• 3-6 CHREC memberships
• 3 memberships allows completion of tasks on
  several scenarios
• Additional memberships expand breadth/number of
  scenarios

16
Conclusions Member Benefits

• Conclusions
• Multilevel algorithm parallelization is critical
  for new apps
• No substitute for exploring, crafting, and
  evaluating
• Build experiential database from hands-on
  experience
• Examining formalized methodologies for RC
  application parallelization will allow more
  efficient development
• Amass and share key insight with apps, design
  methods, and tools necessary are this critical
  stage in fields development
• Concomitant with performance, other apps issues
  often critical
• Fault tolerance, design portability
  standardization, numerical precision
• Member Benefits
• Direct influence over applications, systems,
  languages and FPGA platforms under investigation
• Applicability to current developments and trends
  in aerospace related research
• Access to final code, performance results,
  deliverables