Parallel Paradigms for Numerical Weather Prediction

点击翻译按钮获取中文摘要

Weather forecasts and climate simulations require dedicated highperformance supercomputers to run. Advances in the power ofsupercomputers bring the possibility of simulating the atmosphere athigher resolution (i.e. with more detail) without having to waitlonger for the answer. It has been consistently shown that increasingthe resolution of atmosphere models results in more accurate weatherforecasts and climate simulations. However, getting models that canmake full use of state-of-the-art supercomputers is very challenging.The Met Office is in the process of installing a new Cray XC40supercomputer which which will deliver 16 petaflops (16 quadrillionarithmetic operations per second)peak processingpower by using 4800000 individual processors computingtogether at the same time (in parallel). In the next few decadessupercomputers are expected to deliver more and more computing power, by using more and more processors. The main thing that slows downcomputations on these massively parallel supercomputers iscommunicating data between processors. Unfortunately, the physics of theatmosphere means that the weather in one location is intrinsicallylinked with the weather at all other locations on the globe; thismeans that a lot of data communication between processors is required.Scientists who develop atmosphere models are currently grappling withthe fact that we are close to the limit of what is possible in termsof resolution and simulation speed, due to the communicationrequirements of the mathematical algorithms that are used to solve theequations that predict how the weather evolves in time. At the moment,these algorithms use geographic parallelism: the globe is divided upinto overlapping pieces and each piece is given to a differentprocessor, which must communicate data to processors that sharegeographic locations on the overlaps. To speed up a model, we need touse more and more processors on smaller and smaller regions. Thespeed-up is eventually limited when there are so many overlapping regionsthat all of the globe iscovered by overlaps, and the model spends all of the timecommunicating.This means that it is time to invent new mathematical algorithms thatcan make better use of the parallel computer. In this project we will developalgorithms that are time-parallel as well asgeographic-parallel. Instead of advancing the forecast of the modelforwards step by step in time, these methods produce several differentestimates of the weather at the next step, before combining themtogether to make a more accurate solution. Each of these differentestimates can be independently calculated, which introduces additionalparallel computation into the model.This project is in close partnership with the Met Office. Ifsuccessful, these algorithms will lead to faster and higher resolutionweather forecast and climate prediction models at the Met Office,leading to more accurate forecasts for government, industry and thegeneral public. The Met Office provides forecasts for customers acrossthe transport sector, particularly for aviation planning (so thataeroplanes can avoid headwinds and make use of tailwinds) andpredictions of the motion of volcanic ash clouds. It also providesforecasts for retail and leisure, insurers, the Ministry of Defence, and theEnvironment Agency (including flood forecasting). More accurateforecasts will allow all of these business organisations to plan furtherinto the future, avoiding risks and unnecessary costs.