The abundance of digital data due to the rapid development of ICT has generated entirely new, multidisciplinary approaches in social sciences, in which network theory and computational science play a considerable role both in terms of data analysis and modeling. Nowadays this field is known as computational social science and it has attracted a lot of attention.
In the first half of my talk, I will give a brief overview of computational social science and then introduce our recent researches on modeling social relationships[1,2]. In order to understand the underlying social activities, we propose microscopic models and compare the properties of the generated networks with the empirical observations.
In the latter half, I will introduce two kinds of frameworks for parameter-space explorations which is being developed by our team[3]. One of the most fundamental difficulties of social simulations is that the models for social phenomena are not established well as those for natural sciences. It is also difficult to determine the initial condition or boundary condition even if big data are available since they are usually incomplete. Thus, for social simulation, it is more important to capture the global property of phase space rather than simulating a specific event precisely. To overcome these difficulties, we are developing frameworks which enables us to explore parameter space more efficiently.

参考文献:
[1] Y. Murase et al., "Multilayer weighted social network model", Phys. Rev. E, 90, 052810 (2014) doi: 10.1103/PhysRevE.90.052810
[2] Y. Murase et al., "Modeling the role of relationship fading and breakup in social network formation”, PLoS ONE 10(7) e0133005 (2015) doi: 10.1371/journal.pone.0133005
[3] https://github.com/crest-cassia/oacis
      Y. Murase et al., "A tool for parameter-space exploration", Physics Procedia 57, 73 (2014) doi: 10.1016/j.phpro.2014.08.134

Liquefaction refers to a sudden change in ground behaviour from solid-like to fluid-like. Induced by strong ground motion of an earthquake, soil happens to lose its strength and to start to flow, causing damages to structures on or under the liquefied ground.
Coupling between soil particles and pore waters is a key mechanism of liquefaction. Based on continuum mechanics, governing equations of soil displacement and pore water pressure are derived. Transition from a stable to an unstable solution of the governing equations is regarded as a source of liquefaction initiation. Stability analysis of the solution thus facilitates better numerical treatment of the initiation and development of liquefaction.
We carry out the stability analysis linearizing the governing equations. It is shown that dilatancy (shear deformation inducing volume change) changes the stability of plane waves of soil displacement and pore water pressure; the wave becomes unstable when the dilatancy ratio exceeds a certain critical value [1]. We further carry out numerical analysis that uses K computer for a spherical wave, developing a finite element method based on the particle discretization scheme that is suitable to compute soil particle detachment. It is shown that an unstable solution is captured; the dilatancy effect triggers and the detaching effect spatially expands the unstable solutions [2].
The loss of stability has been overlooked in the numerical analysis of liquefaction. For more rational prediction, therefore, it is necessary to develop a code which accounts for the stability. A finite element analysis that solves this problem is being developed.

参考文献:
[1]. J. Chen, H. O-tani and M. Hori, On Mathematical Stability Analysis of Liquefaction Considering Soil-Water Coupling, Journal of Japan Society of Civil Engineers, vol.70, No.2, I_641-I_648, 2014. doi: 10.2208/jscejam.70.I_641
[2]. J. Chen, H. O-tani and M. Hori, Stability analysis of soil liquefaction using a finite element method based on particle discretization scheme, Computers and Geotechnics, vol. 67, 64–72, 2015. doi: 10.1016/j.compgeo.2015.02.008

Aerosol, which is tiny particle in the atmosphere, affects the microphysical and radiative properties of cloud particles. This effect is called “cloud-aerosol interaction”, which is one of the most uncertain factors for climate predictions. Cloud microphysical model and aerosol transport model, which are parts of component in atmospheric numerical simulation code, have been developed and used to understand the cloud-aerosol interaction. In my talk, I first introduce the cloud-aerosol interaction, the cloud microphysical model, and the aerosol transport model. After that I will introduce several examples about how to use the cloud microphysical model and aerosol transport model for understanding the cloud-aerosol interaction.

One major challenge presented by extreme-scale scientific computing is the huge amount of data that the simulation is capable of generating. Since each run of a state-of-the-art simulation can output data at the petascale, storing all of the data is no longer an option. Aside from simply dropping selected time steps, as has been the practice, additional data reduction methods must be considered in order to meet data movement and storage requirements, while maintaining the accuracy and integrity of the data. This is particularly important as scientific supercomputing moves toward exascale.
Firstly, we proposed a proper orthogonal decomposition (POD) based parallel compression method. A binary load-distributed approach is also proposed for fully utilizing all parallel nodes [1]. However, this method can only deal with the case that both the number of time steps and the number of parallel nodes are power-of-two. In order to resolve this problem, we proposed a 2-3-4 combination method [2]. Furthermore, in order to minimize the error, we proposed an m-swap method [3]. This method allows users to control the compression scale in one node. Recently, we are trying to apply these methods for in-situ visualization.
Besides of the POD-based compression methods mentioned above, a run-length method [4] was also proposed for lossless compression. Now we are proposing a compression method for reducing the size of streamlines and pathlines from large-scale simulations.
These approaches allow us to effectively use all of the processors and to reduce the interprocessor communication cost throughout the parallel compression calculations. The results of tests using the K computer indicate the superior performance of our design and implementation.

参考文献:
[1]. Chongke Bi, Kenji Ono, Kwan-Liu Ma, Haiyuan Wu, and Toshiyuki Imamura, “Proper Orthogonal Decomposition Based Parallel Compression for Visualizing Big Data on the K Computer,” in Proceedings of Eurographics Symposium on Parallel Graphics and Visualization, pp. 1-8, June, 2014.
[2]. Chongke Bi and Kenji Ono, “2-3-4 Combination for Parallel Compression on the K Computer,” in Proceedings of IEEE Pacific Visualization Symposium, pp. 281-285, March, 2014. doi: 10.1109/PacificVis.2014.28
[3]. Chongke Bi, Kenji Ono, and Lu Yang, “Parallel POD Compression of Time-Varying Big Datasets Using m-Swap on the K Computer,” in Proceedings of IEEE International Congress on Big Data, pp. 438-445, June, 2014. doi: 10.1109/BigData.Congress.2014.70
[4]. Shota Ishikawa, Haiyuan Wu, Chongke Bi, Qian Chen, Hirokazu Taki, and Kenji Ono, “Fluid Data Compression and ROI Detection Using Run Length Method,” in Procedia Computer Science, Vol. 35, pp. 1284-1291, 2014. doi: 10.1016/j.procs.2014.08.228

Five years have passed since CUDA, a parallel computing platform and programming environment for GPUs, was released by NVIDIA, the GPU hardware and its software have matured substantially in practical use. Nowadays, GPU computing has been a major topic of research in the HPC area. For linear algebra computations, the GPU acceleration can be widely used not only for compute-intensive operations such as level-3 BLAS, but also for memory-intensive operations such as sparse matrix computation. Although the K computer and the next system do not have accelerators, many core accelerators such as GPUs are still considered as one of the potential technologies for emerging exa-scale supercomputers. In this talk, I will talk about general topics related to the development of linear algebra software on GPUs and some my experiences in this field: (1) overview of recent trends in GPU computing and linear algebra libraries on GPUs, (2) past and future challenges in the development of linear algebra kernels on GPUs, (3) my experience 1: auto-tuning to achieve the best performance, (4) my experience 2: extended-precision arithmetic support on GPUs.

The weather system is chaotic and known to be sensitive to the initial conditions. Therefore, it is important to estimate accurately the current state of the atmosphere for more accurate initial conditions. Data assimilation is a statistical approach to estimate the best possible atmospheric state using both simulation and observation data. The local ensemble transform Kalman filter (LETKF) is an advanced data assimilation method that is particularly efficient with parallel architecture computers.
Non-hydrostatic icosahedral atmospheric model (NICAM) is a weather forecasting model and has been developed by Computational Climate Science Research Team in AICS. Data Assimilation Research Team works collaboratively and applied the LETKF to the NICAM for assimilating various kinds of observations. We are developing a new system for assimilating satellite data. In this talk I will give a brief and clear introduction to data assimilation and present the most recent research on the newly-developed NICAM-LETKF system．