Matematik och Datavetenskap

Energy efficiency has become one of the most important considerations for HPC systems, particularly for large scale systems, for economic and environmental reasons and in exceptional cases also social and political. Many approaches are currently being pursued both in regards to architecture and hardware and software technologies to improve energy efficiency for HPC systems. The prototype described here, one of several within the PRACE project exploring improved energy efficiency, explores energy efficiency achievable through use of commodity components for cost effectiveness, and without acceleration for preservation/ease of portability of the large application code base that exists for the type of HPC systems that have been dominating for a decade. The prototype development was a collaborative effort between industry and academia. With a very limited budget for a server design project and severe time constraints the novelty was effectively limited to careful component choices in regards to energy efficiency for HPC workloads and a new motherboard design to support the component choices. A further constraint was that the outcome would be of production quality in order for the industry partners to market the prototype design should it be successful. For the component choices we did a characterization of the power consumption of a blade chassis and made an effort to measure the energy consumption of different memory modules under HPC workloads, information we could not find neither in the literature nor from memory or system vendors. Memory power consumption in the prototype, as well as most HPC systems, is second only to the CPU, sometimes a close second. We report on the design of the prototype, and preliminary performance results with an emphasis on the energy aspects of benchmarks and compare our results with the Blue Gene/P that, after its introduction, has dominated the top of the Green500 list for systems not using acceleration. The preliminary results show that energy efficiency comparable to the BG/P can be achieved without any proprietary technology at a fraction of the cost. The prototype design is now included in the standard product line of the participating platform vendor.

Math and Computer Science spawned the discipline of computational science which is the bridge that connects observation, theory, and experiment.

Detailed mathematical models simulate physical phenomena from chemical reactions, to the behaviour of biological systems, seismic waves, stars, and even people and financial markets. The value of these models is limited by the available computing power: with greater power, more detailed models lead to more accurate and reliable results.

In global climate modelling, for example, results become more accurate as more subsystems are modelled: the entire atmosphere from troposphere to exosphere; the hydrosphere of oceans, lakes, and rivers; the cryosphere of ice sheets, “While computers use power, they also contribute to energy savings globally.” glaciers and icebergs; the biosphere of animals, plants, and cities; and the solid earth with its mountains, volcanos, and desserts. The application of these coupled models requires computing power that is not yet available in the world’s largest supercomputers.

In the next decade, , supercomputers may increase in power by a factor of a thousand, which will help in addressing the most complex problems facing society, including the development of accurate long term weather predictions, better understanding of climate change, personalized drugs, and predictive health care based on detailed DNA screening.

Scientists are concerned with the correctness, validity, and usefulness of their models, and spend their time using computers as tools to solve their problems. Programmers and computer scientists are concerned with the programmability, portability, efficiency, performance, and getting the most out of the resources available. The bridge between a scientist’s equations and the final application are the algorithms and abstract programming models of computational science. Extensive collaboration among scientists, programmers, and computational science is needed to ensure the best use of the supercomputers of the future. This collaboration should develop domain-specific frameworks and toolboxes for expressing the algorithms and making them more readily port able between different systems. These frameworks will speed up program development, and hide the intricacies of parallelizing computational kernels. Current methods are inadequate to deal with future Exascale systems with millions of cores, especially considering the likelihood of component failure during the execution of a program.

With the development of domain-specific frameworks, scientists in many other disciplines will be able to take advantage of high-performance computing, and the field of computational science will become ever more important.

The recent dominance of multicore processors and the widespread use of heterogeneous architectures have forced a shift towards hybrid programming models. This and the rapidly growing importance of energy efficiency make parallel programming more challenging than in the past. Developing domain specific languages makes it easier for scientists to accomplish programming tasks, but harder to take best advantage of parallelism and diverse architectures. Parallel programming is today more diverse, but also less stable than it was a decade or two ago. Research projects aiming at maximum performance must therefore include computer scientists in joint efforts to address these challenges.