Post on 22-Jan-2018
transcript
The Education of
Computational Scientists
Robert Voigt
Krell Institute
“… computational science has been described as essential to advances throughout society and
deemed one of the most important technical fields of the 21st century.”
https://www.nitrd.gov/pitac/reports/20050609_computational/computational.pdf
Outline
• History
• Formal coursework and programs
• Exposure to multidisciplinary research
• Putting it all together: A successful model
Brief History
When did Computational Science
“Begin”
• Scientific Computing (SC) or Computational Science
and Engineering (CSE)
• 1950s: Von Neumann’s work leading to
– MANIAC
– ORACLE
and their use in solving engineering problems
• Clearly SC/CSE but not recognized as such until
1982 Lax Report
(Report of the Panel on Large-Scale Computing in Science and Engineering,
http://www.pnl.gov/scales/docs/lax_report1982.pdf)
Early View of Scientific Computing
“During its spectacular rise, the computational has joined
the theoretical and the experimental branches of
science, and is rapidly approaching its two older sisters
in importance and intellectual respectability”, Lax, 1986.
Evolution of Computational
Science
O. Yas¸ar, K. S. Rajasethupathy, R. E. Tuzun, R. A. McCoy, and J. Harkin, A new
perspective on computational science education, Comput. Sci. Engrg., 5 (2000)
Computational Science as a
Discipline
Computer
Science
Applied
Discipline
CSE
Mathematics
Yas¸ar, et al. 2000
Petzold et al. 2001
So What is CSE
• Many definitions in the literature
• Key components
– Education
• Mathematics
• Computer science
• Science or Engineering discipline
– Exposure to multidisciplinary research
– Integration of knowledge & methodologies
• More than Computational X
CSE: A dynamic field of its own
• New areas of science and engineering
– From CFD to bioinformatics to social science
– Predictive Science
• Explosion of data
– analysis of data-centric applications
– data-driven scientific discovery and data
enabled uncertainty quantification
– analysis of experiments
• Rapidly changing computer architectures
Formal PhD Programs
• Survey (2012):
http://icl.cs.utk.edu/survey/summary/
• Types of graduate programs:
– A degree in CSE: 41, 19 PhD
– A minor in CSE: 7
– A certificate in CSE: 6
– A track in CSE 2
CSE Course Work
• Most of the 19 have course requirements
involving at least a subset of
– Computer Science
– Applied Mathematics
– Science or Engineering
• What is missing is immersion in
multidisciplinary research
• Many reports cite this as critical to the
development of computational scientists
Immersion in Multidisciplinary
Research: Early Examples• Department of Energy Research Facilities
– Extensive postdoc opportunities
– Seldom combines application, math and CS
• Institute for Computer Applications in
Science and Engineering (ICASE)
– Set inside NASA Langley Research Center
but independent
– Research combining application, math and
CS
– Postdoc opportunities and some
undergraduates
More Recent DOE Examples
• Characteristics
– Partnerships among DOE Labs and Academia
– Applications, math and CS
• ASCR Scientific Discovery through
Advanced Computing (2000)
– Software advances to support applications
• Co-Design (2011)
– Application characteristics influence hardware
and software design
No focus on student development
Academic Example
• Academic Strategic Alliance Program
– Initiated in 1997 by DOE/NNSA Office of
Advanced Simulation and Computing
– Focus on modeling and simulation of
multidisciplinary, multiscale applications
• Predictive Science Academic Alliance
Program (2008)
– Added Verification/Validation & Uncertainty
Quantification research to support prediction
PSAAP II Characteristics
• Funded in 2014 for 5 years at $4M and $2M per
year
• Added specific CS research to support evolution
to useable Exascale computing demonstrated in
the context of their application
• Co-located students immersed in application,
math, CS environment
• Requires internship at NNSA National
Laboratory by all students
• http://www.sandia.gov/psaap/
Stanford UniversityPredictive Simulations of Particle-laden
Turbulence in a Radiation Environment
Use particles to provide higher energy absorption
and transfer rates
• Multidisciplinary– Turbulence, DNS with particles
– Particle transport
– Radiative transport
– Eulerian PDE solver for turbulent flow
– Lagrangian method/particle tracking for solid phase
– Ray tracing/discrete ordinates for radiative transport
• CS research – Domain-Specific Languages (DSL)
• High-level library/framework to dynamically compile
optimized code
• Multiple DSLs for multiphysics coupling
– Containment Domains for resiliency
Exascale Simulation of Plasma-coupled CombustionPredict ignition threshold for jet in crossflow via
spark-discharge and dielectric-barrier-discharge
(DBD) plasmas
• Multidisciplinary– Turbulent mixing and combustion
– Plasma heating, ionization, and transport
(Poisson solver)
– Chemical kinetics and transport
– Prediction of onset of ignition of jet
– Scalable asynchronous numerical algorithms
• CS research – Suite of interoperable tools
– Overdecomposition for locality, latency, load
balancing
– Source-to-source transformations for architecture
opt. while supporting programmer productivity
University of Illinois
Plasma-ignited fuel jet in crossflow
University of Utah
Oxygen-fired “Clean Coal” Boiler for High
Efficiency Electric Power Generation with
Carbon Capture
Predict heat flux for an as yet to be built system
• Multidisciplinary– Turbulence with particle transport
– Radiative transport via rays or discrete ordinates
– Combustion of particles and turbulent gas
– Heterogeneous chemistry of reacting particles
– Higher order numerical methods
• CS research– In-situ analysis and visualization
– DSL for solution of PDE systems
•embedded in C++
•allows for gradual adoption of features, continuous evolution
of large legacy codebases
50-mm coal
particles
High efficiency advanced ultra-supercritical (AUSC) oxy-coal tangentially-fired power boiler
Texas A&M University
Exascale Radiation Transport for High
Energy Density Physics
Validation of photon transport algorithms using
neutrons as surrogates
• Multidisciplinary
– Neutron transport
– Neutron scattering cross-sections
– Adaptive parallel multigrid algorithms
– Scalable iterative methods: Adaptation in time,
space, and angle
• CS research– Multilevel libraries for architecture independence
(like DSLs)
– Methodology for automated selection of optimum
algorithms
– Replication of task graphs for fault tolerance
– Performance models, machine models
Neutron group fluxes
in 1m x 1m x 2m
graphite block, with
pulsed source
Possible 3D
streaming paths in
neutron transport
experiments
University of FloridaCompressible Multiphase
Turbulence in Explosive-driven
Particle-laden Flows
Multiscale models for shock/particle interaction
• Multidiciplinary– Turbulence with particle interaction
– DNS and LES
– Hybrid spectral-WENO schemes
– High-order discontinuous Galerkin methods
– Heterogeneous space-time discretizations• Perform computation only when and where needed
• CS research
– Field-programmable gate array (FPGA)
emulation at device, node, and system-level
– Behavioral Emulation Objects representing
software/hardware for performance prediction
Explosive dispersal of 114-mm Al particles
University of Notre Dame
Shock Wave-processing of Advanced Reactive Materials
Predict conditions for synthesis of cubic boron
nitride (c-BN) by reverse Taylor impact exp
• Multidisciplinary– Shocks in heterogeneous reactive materials– Solid-solid phase transitions, plastic flow, shear band formation, grain sliding, debonding– Micro/meso/macro scale bridging– Simulation of shock-induced synthesis of Ni/Al
composite
• CS research– Asynchronous parallel computing model
– Dynamic adaptive control of computing and resource management
• Light-weight user threads
• “Move the work to the data” when advantageous
• Light-weight semantically rich synchronization
mechanisms like dataflowStress in a Ni/Al composite
Training and Education
• Opportunities for training via national
programs
– NSF and DOE Supercomputer Centers
– NSF Extreme Science and Engineering
Discovery Environment
– NSF Research Traineeship Program
– DOE National Laboratories
• Informal education in CSE that fills in
missing gaps in academic programs
DOE Computational Science
Graduate Fellowship (CSGF)
• Administered by the Krell Institute
https://www.krellinst.org
– With support from the DOE
– Guidance from Steering Committee
• Undergraduates & 1st year graduates
eligible
• Students complete fellowship application
on line focused on a science/engineering
application
Program Support
• Stipends ($36,000/year for 4 years)
• Full tuition and fees
• Professional development support
• $5,000 first year and $1,000 each renewed
year
• Laptop/conference travel/society dues …
• Practicum support (living expenses and travel)
• Annual program review
• Fellows “own” the fellowship
CSGF
• Includes three major components
– Formal education
– Focus on an application
– Immersion in DOE Lab via a 12 week practicum
• Application requires essays & courses
• Complex, thorough review of applications &
renewal process
• Continual monitoring of progress
Formal Education
Program of Study (POS)• Proposed POS must contain
– Year of courses in the chosen application
– Year of courses in math/statistics
– Year of courses in computer science
• Courses must be from traditional
department for the discipline
• POS must contain exposure to HPC
• POS is continually reviewed by Steering
Committee
Practicum
• Each fellow required to spend 12 weeks at
a DOE facility
• Fellow submits a proposal reviewed by
Steering Committee
• Immersed in a multidisciplinary research
environment
• Mentored by active researcher
• Intended as a broadening experience
Some Statistics
• CSGF entering its 26th year
• Approximately 400 alums
• Annual awards range from 5 to 23
– Driven by budget and federal policy
– Approximately $10M annual budget
• Highly competitive
– 456 applications for ~ 20 awards this year
• Large research universities dominate
Where do Alums End UP
Krell has current employment data for 314 alumni grouped as follows:
– Academia 101
– Industry 105
– DOE Labs 47
– Other Gov’t. 28
– Grad. Student 25
– Other 4
– Nonprofit 4
– Unknown 15
Academia31%
DOE Labs14%
Graduate Student
8%
Industry32%
Other1%
Other Government
8%
Unknown5%
Non-Profit1%
Alumni by Employment Category
0
2
4
6
8
10
12
14
16
18
1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Incoming Fellow Classes by Field Area
Biology and Bioengineering Computer Science and Math Engineering Physical Science
$-
$2,000,000
$4,000,000
$6,000,000
$8,000,000
$10,000,000
$12,000,000
$14,000,000
FY2004 FY2005 FY2006 FY2007 FY2008 FY2009 FY2010 FY2011 FY2012 FY2013 FY2014 FY2015 FY2016
DOE CSGF Funding (FY2004-FY2016)
ASCR NNSA Total
16 15 1917
18 16
21 18
11
21
10
23
27
Funding Challenge
What’s Missing in CSGF
• Present program focus on application is
barrier for Math/CS majors
• Add enabling technologies track
– Focus on HPC
• Exascale algorithms
• Data analytics
• Machine Learning
– Still require exposure to applications but not
the focus