Dynamic Exascale Global Address Space Programming Environments
The Dynamic, Exascale Global Address Space programming environment (DEGAS) project will develop the next generation of programming models, runtime systems and tools to meet the challenges of Exascale systems. We will develop a new set of programming concepts based on a hierarchical model of parallelism and data locality, hierarchical fault containment/recovery for resilience, introspective dynamic resource management, demonstrate them using extensions to existing languages, and evaluate their utility for applications. Our solution will address the following key challenges posed by exascale systems:
- Scalability: Efficient communication (extended GASNet) and synchronization mechanisms combined with compiler (ROSE) and runtime optimizations to minimize both.
- Programmability: Rich set of programming constructs based on a dynamic, resilient Partitioned Global Address Space (PGAS) model, demonstrated in multiple language dialects (C and FORTRAN).
- Performance Portability: Non-invasive profiling (IPM), deep code analysis (ROSE) and a dynamically Adaptive RunTime System (ARTS).
- Resilience: Containment Domains and state capture mechanisms and lightweight, asynchronous recovery mechanisms.
- Energy Efficiency: Runtime energy adaptation and communication-optimal code generation.
- Interoperability: Runtime and language interoperability with MPI, OpenMP and libraries (Lithe).
The DEGAS team will work with Co-Design centers to drive the programming construct design, combined with information about hardware platforms as it emerges. We will also leverage ongoing discussions with other application and vendor stakeholders as well as mainstream language standards groups, augmented with advisory committees and semi-annual retreats involving broad representation from all three groups.
Our approach focuses on a vertically integrated programming and execution environment that incorporates the latest algorithmic approaches and application structures to effectively service ultra-scale science and energy applications. The primary focus areas of DEGAS are shown in Figure 1 along with the proposed integrated software stack.
- Katherine Yelick, Principal Investigator (LBNL)
- Krste Asanović (UC Berkeley)
- James Demmel (UC Berkeley)
- Mattan Erez (UT Austin)
- Paul Hargrove (LBNL)
- Steven Hofmeyr (LBNL)
- Costin Iancu (LBNL)
- Khaled Ibrahim (LBNL)
- John Mellor-Crummey (Rice University)
- Leonid Oliker (LBNL)
- Dan Quinlan (LLNL)
- Eric Roman (LBNL)
- Vivek Sarkar (Rice University)
- Erich Strohmaier (LBNL)
- Yili Zheng (LBNL)