HPC University RAT

Overview Prepared forTeraGrid Quarterly Meeting

Austin, TXDecember 6-7, 2007

Introduction – What are we working on?

HPC University RAT CharterThe HPC Training RAT will identify successful paths to

creating qualified, effective HPC professionals, capable of exploiting current super-terascale and upcoming

petascale technologies for the advancement of scientific research. Gaps in training materials and delivery

methods will be identified, and recommendations for filling these gaps will be presented.

MembersBonnie Bracey-Sutton Diglio SimoniJohn Cobb Stephenie McLeanShawn Brown Christian ChilanBrad Armosky Sudhakar PamidighatamJames Rome John ZiebarthMary Ann Leung Scott LathropKay Hunt Subha Sivagnanam

Krishna Madhavan

AdvisoryAlina Bejan

Steve GordonLeslie Southern

Robert PanoffKristina Wanous

RAT Report

•Catalog Map– Topics– Mastery Levels

•Quality Assurance•Delivery Methodologies•Target Populations

– Demographics– Disciplines

•Scaling & Dissemination•Getting to Petascale

Catalog Map

NoviceUndergrad

ApprenticeMaster

JourneymanPh.D.

MasterPost-doc/Prof

Modeling/ simulation

HPC Technology (hardware)

DomainsSoftware engineering

Operational issues

ArchitecturesPerformance analysis

Scalable computing

Application packages

Programming/ Algorithms

Code optimization

Science gateways

Development tools

Workflow management

Visualization

Data analysis/ Post-processing

Verification/ validation

Collaboration

At least 10 offeringsLess than 10 offeringsNo offerings

Catalog Map – Recommendations

•Complete the verification and validation process for all entries in the catalog

•Develop a mechanism for the continual update of the catalog to identify new sources of training materials

•Make the catalog readily available to the HPC community in a persistent way

Quality Assurance

•CSERD/NSDL Overview– Digital repository for instructional & reference materials for

computational science resources– Includes comprehensive review mechanism for materials,

similar to journal review process– “Solving the problem correctly” != “Solving the correct

problem?”

•Verification– the resource works as advertised on the computing

platforms as advertised

•Validation– the resource is based on current, valid scientific methods

•Accreditation– the resource is appropriate for the advertised audience

Quality Assurance – Recommendations

•Leverage CSERD/NSDL:– Broad dissemination– Formal VV&A review process.

•Resource Management; i.e. identify a “primary editor”– Finding reviewers, – Organizing reviews– Publishing the resources that have passed a minimum

quality level

•Broadening Community Involvement– Contribute additional materials– Offer community comments on the resources– Provide feedback on the resources to TeraGrid on a regular

basis

Delivery Methodologies

TopicNumber

of EntriesDelivery Method

Asynchronous

Synchronous

Face-to-Face

Operational Issues 60 Programming & Algorithms 40 Development Tools 37 Architectures (Parallel, Dist, Grid) 21 Science gateways & resources 18 Performance Analysis 12 Visualization 10 Data Considerations 8 Scalable Computing 7 HPC Application packages 7Collaboration 5 Modeling & Simulation 4 Code Optimization 4 HPC Technology (hardware) 3 Domains (physics, chem, etc.) 2 Other 1 Workflow management 0Software Engineering 0Verification & Validation 0Data analysis/post processing 0

This is a holeTotal number of offerings: 239

Delivery Methodologies – Recommendations

• Plug the holes - provide a variety of teaching methods for each topic area– Allow students to customize training programs to meet their

learning styles.• Redesign course materials to produce effective interactive online

learning experiences– Avoid simply putting lecture-based materials online

• Utilize efficient multi-disciplinary teams, including – subject matter experts– instructional designers & technologists– information technologists– tutors– faculty– project managers

• Incorporate metacognitive* skills development into all training methods – Create or adopt a module that teaches students how to become

aware of their most effective learning styles– Supplement these materials with guidelines on how to select the

appropriate training materials based on their learning styles

Population Analysis

•Known problems– Poorly represented demographic groups– Scientific disciplines that are not using HPC– Costs of delivery, esp.

•To economically challenged populations•In temporally effective ways

•In particular, we don’t have the data to know what the population issues are– TG Core Services is addressing some of this in the

Core2 project

Population Analysis – Recommendations

•Proactively support efforts to Broaden Participation across– Underrepresented demographics– Underrepresented fields of science

•Broaden delivery methods– Synchronous: eliminates geographic boundaries– Asynchronous: eliminates temporal boundaries

Scaling & Dissemination

•Train-the-trainer– Most people teach the way they were taught– BUT, most survey respondents are self-taught– So…how do we scale and disseminate HPC training?

•Collaboration & Mentoring– Time-intensive for “master”, but effective methodology

(long history of success)– Traditionally geographically & temporally constrained

•Technology-du-jour - “Masters” may be disinclined to learn new technology– Learning curves– Versioning & dialects– “Pain threshold” for dealing with reliability issues

Scaling & Dissemination –Recommendations

• Identify and promote Best Practices for content development and delivery

•Encourage mentoring within TeraGrid– Create a Web portal for people to volunteer their services– Create a Web Portal for users to ask questions and to get a

mentor– Provide real incentives for such volunteer activity– Study the social interactions in mentoring situations to

provide training to make mentoring more effective•Provide support to HPC trainers

– Training accounts on TeraGrid– Help with developing and critiquing course content

•Provide a Web Portal for users new to HPC that would provide a roadmap of training that would lead them to their individual goals

Getting to PetaScale –Recommendations

•Engage scaling & PetaApps experts to suggest directions for these issues

•Collaborate with Petascale System vendors to provide tools that will make it easier to use their machines

•Include petascale applications and scaling experiences in the case study libraries being developed as TG EOT initiatives

Low-Hanging Fruit

•Support Pathways (EOT Supplement) to broaden participation– Case Studies– Campus Champions– TG Fellows

•Workshop on “How to Teach HPC” (train the trainers)

•Develop AG Training framework for well-established existing material– How to Write a TG proposal– Welcome to the TeraGrid

Longer-term Initiatives

•Establish resource archive – make training materials persistent– Can’t just take face-to-face and make it on-line

training

•Establish mentoring program– Identify mentors who can make the commitment

•Identify reviewers who can verify/validate existing offerings– Begin VVA process