12.5km & 6km SAMIL3 is successfully operated on the "Tianhe" 10000-core Supercomputer
SAMIL3, the latest in the series of global atmospheric general circulation models developed at the State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics (IAP), successfully achieved two groups of simulations with more than 10,000 cores on “Tianhe-1A” supercomputer platform at National Supercomputer Center in Tianjin, China. SAMIL3 has been tested with 12.5km and 6.25km respectively in horizontal resolutions. That is the first attempt in China to perform the simulation of atmospheric general circulation model on more than 10,000 cores supercomputer.
The dynamic core of SAMIL3 contains the finite volume method (Lin, 2004) and the spectral method (Wu et al., 1996). SAMIL3 can be operated with various resolutions, including 200km, 100km, 50km, 25km, 12.5km and 6.25km. Meanwhile, the number of I/O nodes is also changeable. Based on these, SAMIL3 has been tested with large amounts of standard dynamical core standard cases and Aqua Planet Experiments (APE).
APE applies AGCMs with their complete parameterization package to an idealization of the planet Earth that has a greatly simplified lower boundary that consists of an ocean only. It has no land and its associated orography, and no sea ice. The ocean is represented by Sea Surface Temperatures (SST) which are specified everywhere with simple, idealized distributions.
Model speed experiments (figure 1) have demonstrated that SAMIL3 has excellent scalability under 1000 cores, which relative parallelized efficiency lies between 1.0 and 1.14. As core number increases, relative parallelized efficiency keeps going without significant loss until using more than 3456 (12.5km) and 6144 (6.25km) cores when they drop to 0.66 (12.5km) and 0.52 (6.25km). That is to say SAMIL3 has the best scalability under 3456 (12.5km) and 6144 (6.25km) cores, at which the model can reach about 1.5 model years (12.5km) and 3.5 model months (6.25km) per wall clock day.
In model I/O efficiency experiments, (figure 2) zonal wind (u), meridional wind (v), specific humidity (q), air temperature (t) and surface pressure (ps) are output at each time step. Among them, u, v, q, t are three-dimensional variables, with 26 levels on the third dimension, and ps is a two-dimensional variable. Roundly 200GB data are generated in each case of this group of experiments. The results demonstrate that if there is no I/O, I/O operations at a free cost, and the efficiency is 100%. When 6 I/Os are opened, model I/O efficiency maintains no less than 80%. As I/Os continue to increase, model I/O efficiency becomes higher. If all cores have I/Os (1536 cores), model I/O efficiency approaches 100%, indicating that SAMIL3 has high I/O efficiency.
Development of high resolution climate model is nowadays the most cutting-edge research of climate simulation around the world. Along with the previous successful experiments of 10km resolution LICOM on 1,000+ cores, accomplishment of 10,000+ core scale simulation of SAMIL3 indicates that our climate model has the capability of competing with other high resolution model research groups, and lays the foundation of ultra-high resolution climate model development in China.
Preliminary simulation results of SAMIL3 on APE can be found at
Figure 1 Model speed for 12.5km (upper panel) and 6.25km (bottom panel) resolution SAMIL3 on APE as a function of Tianhe-1A supercomputer core number.
Figure 2 Model I/O efficiency for 12.5km resolution SAMIL3 on APE using 1536 Tianhe-1A supercomputer cores as a function of I/O numbers.