What can the power of an accelerator processor achieve? A lot “more”, say scientists at the Indian Institute of Science (IISc), Bangalore. In a recent interdisciplinary study involving researchers from the Centre for Atmospheric and Oceanic Sciences and the Department of Computational and Data Sciences, the researchers have used the Intel Xeon Phi coprocessor to speed up intensive calculations in climate modelling. Intel’s Xeon Phi is a series of computer processors which has multiple cores and is designed for supercomputing.
Climate modelling is the process of developing mathematical and computational models using factors like wind speed, latitudes, longitudes, temperature, humidity, atmospheric pressure, etc. of a region and predicting the climate of that region for centuries to come. This process involves huge calculations and is so tedious that even high-speed computers take several days, sometimes months, to complete them. Most climate models are represented as mathematical equations and incorporate the laws of physics such as conservation of mass, energy and momentum. Using the present values of atmospheric factors, these models predicts the future climate and related changes.
“Our research develops performance optimization techniques for speeding up climate simulations with state-of-art accelerators like Intel Xeon Phi”, says Prof. Sathish Vadhiyar, the corresponding author of the study from the Department of Computational and Data Sciences of IISc.
In this study, the researchers have used Community Earth System Model (CESM), a mathematical representation of earth consisting of modules relating to atmosphere, ocean, ice, land surface, carbon cycle, etc. It is developed and maintained by the National Centre for Atmospheric Research, USA. The model is so complex that when simulations are run on CESM, the atmospheric module itself takes 85% of the computational time to complete. The researchers have tried to leverage the developments in computer hardware by running CESM on the Intel Xeon Phi processor.
The research has developed two modes of computation - synchronous and asynchronous. In the synchronous mode, the CPU waits for the result from Intel Xeon Phi coprocessor and uses it for the next computation, while in the asynchronous mode the CPU immediately proceeds to the next computation without waiting for the result of the current computation. The researchers wanted to find the best configuration for running the CESM simulations.
The researchers have also incorporated the techniques of optimization of computer algorithms for faster climatic simulations, thus easing the life of meteorologists from the hassle of computer science. The use of optimization techniques like vectorization (dividing a complex block of calculations into smaller independent blocks, which can run on the separate cores of the processor) and replacement of time-consuming mathematical function by simple interpolation functions have improved the performance of the code by 43% to 55%, claim the researchers.
This study is the first of its kind to compare different models of executions of a climate model on Intel Xeon Phi coprocessors. The researchers found that the asynchronous mode results in faster computation than the synchronous mode and the latter results in better accuracy. They also analysed the performance - accuracy tradeoffs of these modes and have observed a performance improvement of 45% in case of asynchronous modes. This improvement in performance translates to a net saving of two years for multi-century climate prediction!
“Our research can help in multi-fold increase in simulation throughput of climate modelling applications. Thus, climate scientists can perform more exploratory studies in limited time, focusing more on the scientific problems in hand rather than on the time taken for the simulations to complete on computers”, remarks Prof. Vadhiyar.
The applications of this research are manifold. “ Our techniques are generic and can be applied to legacy scientific applications of different domains including molecular dynamics, stencil computations, adaptive mesh refinements and cosmology simulations”, signs off Prof. Vadhiyar.