Energy-efficient HPC through better algorithms
‘Current HPC platforms consume such a huge amount of energy that the further successful development and use of these systems will strongly depend on optimisation towards energy efficiency,’ says Prof Vincent Heuveline, coordinator of EXA2GREEN at the University of Heidelberg. It is this precise thought that led him and his team to initiate the project, with a singular approach. Whilst the team recognises that advances in hardware design and manufacturing will lead to a significant increase in energy efficiency, they decided to focus on other, untapped potential: that of algorithmic design and software engineering. The three-year project ended in October 2015. It resulted in a software tool able to trace and analyse the power and energy consumption of parallel scientific applications. But in order to get there, the team first had to define what makes an energy-efficient algorithm. ‘The first step towards an energy-efficient algorithm is to analyse its behaviour,’ Prof Heuveline notes. ‘How much energy does it consume overall, where is most of it spent, where are the bottlenecks, what is the power draw in each part of the algorithm, and are there any inefficient parts?’ He continues: ‘To conduct such analysis, measurement devices, tracing tools, as well as integrated profiling and inspection tools are needed. The knowledge gained can then be used to improve the inefficient parts of algorithms. Typical measures include the elimination of busy waiting, optimisation of task scheduling to better exploit the available hardware, or even adapting to other hardware which better fits the needs of a specific algorithm.’ The team developed various numerical algorithms, particularly on linear algebra routines and solvers which play an important role in many scientific applications. Software packages were developed that can be used by researchers or deployed on HPC infrastructure, so that the project’s results have an immediate effect on the energy efficiency of all applications using them. Bringing research to fruition: COSMO-ART weather forecast model As any scientific idea or method requires a proof of concept, another key task for the EXA2GREEN team consisted in identifying a suitable model system. They opted for COSMO-ART – a software solution based on the operational weather forecast model of the COSMO consortium that enables the calculation of gases and aerosols’ interaction within the atmosphere. ‘It was a deliberate choice to use an application whose optimisation would have tangible benefits for this important research community,’ Prof Heuveline explains. ‘We have used several complementary techniques to reduce the energy consumption of the COSMO-ART numerical weather prediction model. For example, we investigated the use of different floating point precision formats, improved the methods for computing the temporal evolution of the atmospheric chemistry, used advanced parallelisation techniques and ported the whole application to new hardware platforms.’ Eventually the team managed to divide by three the overall energy consumption of COSMO-ART whilst also achieving a factor four speedup of the simulation. ArduPower: the power inefficiency tracker Although the project, when conceived, wasn’t oriented towards the commercial exploitation of its developments, another of its key contributions is ArduPower – a new type of internal power meter which can be used to measure the power consumption of individual hardware components like computer processor or memory. ‘The development of such a new measurement device was important to complete the algorithm analysis framework, since there were no adequate solutions available before,’ Prof Heuveline elaborates. He says that ArduPower might be a promising candidate for creating a commercial product or service. Since the project was completed, the team has continued working in the field of energy-efficient HPC. They are active in the ICT-Energy collaboration and have been promoting energy-efficiency in large scale applications. In November 2015, one of the project partners, IBM Research Zurich, won the Gordon Bell Prize for its energy-scalable simulation of Earth's mantle convection run on 1.6 millions of BlueGene/Q cores.
Keywords
ArduPower, energy efficiency, algorithms, software, HPC, high performance computing, weathercast, COSMO-ART, EXA2GREEN
