August 2013

HPCG: A New Performance Metric

Since 1993, ICL’s Jack Dongarra has maintained a ranking of the world’s top performing supercomputers. The semiannual TOP500 list is compiled using ICL’s High Performance Linpack benmark (HPL). HPL and the TOP500 are the most widely recognized and discussed metrics for ranking HPC systems. However, the Linpack benchmark, which relies on solving dense linear equations to calculate performance, is becoming less relevant as it does not depict the real-world performance of modern HPC applications, which are moving towards differential equations.

Very aware of this performance ‘gap,’ Jack teamed up with Michael Heroux from Sandia National Laboratories to develop the High Performance Conjugate Gradient (HPCG) benchmark, which—because it is composed of computations and data access patterns more commonly found in applications—will hopefully allow for a better correlation to real scientific application performance and drive computer system design and implementation in directions that will better impact performance improvement.

Linpack will not be retired, however. Instead, HPCG will serve as an alternative ranking of the TOP500 list, allowing a re-ranking of the systems on the list to “real” applications—not unlike the re-shuffling of the list done for the Green500. Dongarra hopes to debut the HPCG benchmark at SC13.

News of PaRSEC Award Spreads

As reported last month, DOE awarded funding to ICL for the Parallel Runtime Scheduling and Execution Controller (PaRSEC) project. The award, which allocates ICL/UTK $1 million dollars over three years, was enough to grab the attention of HPC Wire (by way of TN Today). Congratulations once again!

The Parallel Runtime Scheduling and Execution Controller (PaRSEC) is a generic framework for architecture aware scheduling and management of micro-tasks on distributed many-core heterogeneous architectures. Applications we consider can be expressed as a Direct Acyclic Graph of tasks with labeled edges designating data dependencies. DAGs are represented in a compact problem-size independent format that can be queried on-demand to discover data dependencies in a totally distributed fashion.

PaRSEC assigns computation threads to the cores, overlaps communications and computations and uses a dynamic, fully-distributed scheduler based on architectural features such as NUMA nodes and algorithmic features such as data reuse. The framework includes libraries, a runtime system, and development tools to help application developers tackle the difficult task of porting their applications to highly heterogeneous and diverse environments.

SILAS Project Funded

The Sustained Innovation for Linear Algebra Software (SILAS) project, a collaborative effort between UTK, U of C Denver, and UC Berkeley, was recently funded by an NSF SSI call. SILAS updates two of the most widely used numerical libraries in the history of computational science and engineering—LAPACK and ScaLAPACK, (abbreviated Sca/LAPACK)—enhancing and hardening them for this ongoing revolution in processor architecture and system design.

The primary impact of SILAS is a direct function of the importance of the Sca/LAPACK libraries to many branches of computational science. The Sca/LAPACK libraries are the community standard for dense linear algebra and have been adopted and/or supported by a large community of users, computing centers, and HPC vendors.

Application domains where the Sca/LAPACK libraries are used include (among a host of other examples) airplane wing design, radar cross-section studies, flow around ships and other off-shore constructions, diffusion of solid bodies in a liquid, noise reduction, and diffusion of light through small particles. Enhancing these libraries through the SILAS project with state-of-the-art methods and algorithms, and adapting them for new and emerging platforms, will have a large impact on the research and education community, government laboratories, and private industry.

ADAPT Project Funded

The ADAPT project, recently funded under an NSF SSE call, enhances, hardens, and modernizes the Open MPI library and creates a viable foundation for a new generation of Open MPI components by implementing fundamental software techniques that can be used in many-core systems to efficiently execute MPI-based applications, and to tolerate fail-stop process failures, at scales ranging from current large systems to the extreme scale systems that are coming soon.

To improve the efficiency of Open MPI, ADAPT integrates, as a core component, knowledge about the hardware architecture, and allows all layers of the software stack full access to this information. Process placement, distributed topologies, file accesses, point-to-point and collective communications can then adapt to such topological information, providing more portability.

The ADAPT team, which includes UTK and the University of Houston, is also updating the current collective communication layer to allow for a task-based collective description contained at a group-level, which in turn adjusts to the intra and inter-node topology. Planned expansion of the current code with resilient capabilities allows Open MPI to efficiently survive hard and soft error types of failures. These capabilities can be used as building blocks for all currently active fault tolerance proposals in the MPI standard body.