Wei Chen
CS267 HW0 Fall 2002
The Earth Simulator Project
The Problem
Launched in Japan in 1997, the Earth Simulator (ES) project [1] seeks to create a "virtual earth" to achieve the realization of the prediction global-scale phenomena that greatly influence our daily lives. With an accurate climate model of the earth, scientists will be able to understand the patterns of global climate change, thereby allowing us to anticipate and respond more effectively to the occurrences of natural disasters. For example, one of the project's major goals is to understand the effects of global warming in the hope of finding ways to mitigate its damages on future climates [2]. Other research areas of the project include weather forecast, prediction of typhoon routes, and the occurrences of El Nino. The project's simulation of the earth's environment benefits not only meteorology but other fields in earth science as well; another goal of the project is to advance the understanding of long range crustal movements and the mechanisms of seismicity.
Application of the ES project: prediction of annual precipitation
Applications of Parallelism
Predicting the future is a risky business, especially so when the target of the prediction is as complicated as earth itself. The scale and complexity of global climate changes forbid direct experiments on such phenomena, and computer simulations thus becomes the most viable approach for studying them. Because of the sheer amount of data associated with these meteorological events and their inherent parallel nature, parallel computing is used to perform the simulations; for weather simulations, the atmosphere is divided into grids and assigned to different processors, and the physical quantities such as temperature and humidity can now be calculated in parallel at each individual grid point. With this multi-grid method, to accurately model the environments one must have a high grid resolution. Smaller grid sizes, however, implies longer simulation times that may not be feasible in practice, and current climate models have their resolutions limited to the 100 km range, too coarse-grained for representing narrower cloud systems [3]. One goal of the ES project is to eliminate this source of inaccuracy by reducing the grid size to the 10km range; to achieve this, a large-scale supercomputer called the Earth Simulator is built to speed up the simulations.
Smaller Grid Size Means More Accurate Model
The Earth Simulator
The Earth Simulator, located in Yokohoma, Japan, sits at the top of the current Top 500 [4] list with a peak performance of 40 TeraFLOPS per second and a LINPACK benchmark peak of 35.6 TFLOPS/second. Following is the performance spec of the Earth Simulator [5]:
In other words, ES's superior speed can be attributed to all of the following: fast CPUs, fast memories and fast interconnect switches. The machine runs a UNIX based operating system with support for a parallel programming environment, and claims to scale well up to 640 nodes [6]. As the world's fastest supercomputer, it's almost 5 times as fast as current second place holder on the Top 500 list. Its dominance in performance is further reflected by estimates that the ES may be faster than the sum of the first 19 other machines on the same list.
Model of the ES (approx. dimension: 71 * 55 yard)
Results and Evaluation
Weather simulations executed on the Earth Simulator have offered very promising performance numbers. AFES, a spectral-method atmospheric general circulation model optimized for the ES platform, is able to achieve a sustained performance of 26 TFLOPS/sec (65% of peak) when all 5120 nodes are used. Moreover, AFES has achieved a resolution of 20km grids, which is a dramatic improvement over the standard 300km grid size. Thus, by effectively utilizing the immense computational power of the massively parallel Earth Simulator, scientists are able to simulate complex natural phenomena with high resolution and come up with more accurate predictions of global climate changes and meteorological disasters.
References
1. Outline of the Earth Simulator
2. Toward more Accurate Global Warming Prediction Using the Earth Simulator
3. A management for enormous data output by AGCM (AFES) on the Earth Simulator