Models without Borders

DEPARTMENT OF COMPUTER SCIENCE @ LOUISIANA STATE UNIVERSITY

Models without Borders

Thomas Sterling Arnaud & Edwards Professor, Department of Computer ScienceAdjunct Faculty, Department of Electrical and Computer EngineeringLouisiana State University

Distinguished Visiting Scientist, Oak Ridge National LaboratoryCSRI Fellow, Sandia National Laboratory

April 5, 2011

Presentation to: PEMWS-2

The Status Quo

DEPARTMENT OF COMPUTER SCIENCE @ LOUISIANA STATE UNIVERSITY 2

… or Maybe Not


The Challenge• Enable Extreme scale computing in this decade

– Exascale– Strong scale– Revolutionize STEM and national defense applications– Provide practical, programmable, dependable systems

• Respond to disruptive technology drivers– Multi/many core– Heterogeneous GPU– Flat-lined individual core performance

• Guide future system codesign– Programming models– Operating system and runtime software stack– System and core architectures


The Strategy• Derive a new execution model

– Unified to address all challenges– Single to build community wide adoption– Open to facilitate commercial products and application

• Develop proof-of-concept reference implementations– Developing a diversity of software realizations and tools– Targeting a diversity of present and future architectures– Spin-off early products for HPC adoption

• Empirical studies for quantitative evaluation– Driven by STEM and security related applications– Comprehensive for confidence in capability and reliability


Strategic Requirements• Performance

– Efficiency– Scalability

• Energy– Bounded power– Minimized energy

• Reliability– Continued operation in the presence of faults

• Programmability– System transparency– Portability across system classes, scales, and generations

• Generality– STEM– Knowledge


Tactical Performance Requirements• Starvation

– Insufficiency of concurrency of work– Impacts scalability and latency hiding– Effects programmability

• Latency– Time measured distance for remote access and services– Impacts efficiency

• Overhead– Critical time additional work to manage tasks & resources– Impacts efficiency and granularity for scalability

• Waiting for contention resolution– Delays due to simultaneous access requests to shared physical or

logical resources


PEMWS-2 Objectives• Inform and exchange

– Share experiences and perspectives towards a common goal

• Build a ParalleX community– Bring together sources of complementing ideas– Identify stakeholders in government, industry, & academia– Define common needs and applications

• Establish a path forward– Towards a single unified execution model– Responsive to needs of industry, national agenda &

missions


Major Topics• Execution models as a service to the future• Codelets• Memory models• Parcels for message-driven computation• XPI – towards a low-level ParalleX API• SWARM – many core operation• Habanero – parallel programming environments• Self-aware declarative control• HAD-HPX results• Dialog and perspectives


A Quick ParalleX Review• Synchronous Domains• AGAS – Active Global Address Space• ParalleX Processes – with capabilities protection• Computational Complexes – threads & fine grain dataflow• Local Control Objects – synchronization and global distributed

control state• Distributed control operation – global mutable data structures• Parcels – message-driven execution and continuation

migration• Percolation – heterogeneous control• Micro-checkpointing – compute-validate-commit• Self-aware – introspection and declarative management


Progress Report• ParalleX execution model

– Report version 1.11b• HPX-3

– Targeted to conventional platforms– Missing: processes, scalable AGAS

• Applications– Adaptive Mesh Refinement– N-body Barnes Hut– PXGL– Linpack

• XPI


Models without Borders (1)• Institutional

– Not owned or claimed by any one closed organization– Not proprietary

• Open source– Accessible by all academic researchers– Basis for commercial deployment and deliver vehicle

• Target platform– Portable across classes, scales, and generations– Exploits unique strengths while not limited to them

• Applications– Returns parallel processing to broad range of problems– Array and graph, static and dynamic, numeric and symbolic– Data scaled and strong scaled


Models without Borders (2)• Inter-agency

– Shared responsibility and credit for development• International

– Recognized across continents– World-wide market for compliant products

• Challenge set– Robust under varied technology variations– Resilient in performance across diverse structures

• Convention– Not limited by status quo– Vested interests in restricting change


Models without Borders – Requirements• Federation of Interested Organizations• Charter and Mission of goals and process• Stable funding• Metrics of success• Paradigm enabled – challenge driven• VMG – velocity made good

– Measured progress• Reference implementation

– open to all, Contributed by many– Deliverable schedule of usable components– Modular architecture for distributed development


Conclusions• HPC has a narrow window of opportunity as it

gropes for new balance point in the face of technology challenges

• As before a change in execution model is essential for useful progress toward capable scalable systems

• ParalleX is a strong contender for this new model• ParalleX provides a foundation for development of a

consistent and comprehensive model• Intel/UDel bring needed expertise and concepts for

achieving enhancements critical to ParalleX success• Carpe Diem


Models without Borders

Documents

louisiana state university3

louisiana state university2

department of electrical

system classes

academiadefine common

remote access

parallex communitybring

future architecturesspin