We present our work on developing a unified simulation framework for efficient computation of time resolved approximations for complex industrial flow problems. To address the challenges of modern and emerging supercomputers, efficient data structures and communication patterns are needed. Here, we use a Cartesian grid together with a Lagrangian based immersed boundary method to accurately capture moving, complex geometries. The asymmetric workload of the immersed boundary is balanced by a predictive dynamic load balancer, and a multithreaded halo-exchange algorithm is employed to efficiently overlap communication with computations. Our work also concerns efficient methods for handling the large amount of data produced by large-scale flow simulations, such as scalable parallel I/O, data compression and in-situ processing.

Current trends in processor architecture is increasing the number of cores. Modern CPUs now have 2~16 cores and the Intel Xeon Phi coprocessor has more than 60 cores with 4 hardware threads. This trends force the user to describe fined-grained task-parallelism to exploit lots of cores within a chip. The current programming models in High Performance Computing area lack the ability of describing fine-grained tasking. Combining them with thread-level programming model such as OpenMP is not sufficient because task dependency requires data movement which may cause inter-node communication. The aim of the research is to design and implement programming interface for fine-grain task-parallelism in a PGAS language, named XcalableMP. XcalableMP supports data-parallelism among nodes by describing directives to the serial code. The extended task syntax may include inter-node communication to data exchange between tasks as well as execution dependencies among tasks. The presentation will show the prototype implementation of task-parallelism and performance evaluation of the task parallelism in the XcalableMP framework.

This talk presents a scalable method for exposing and exploiting hidden localities in production stencil applications. An end-to-end framework automatically transforms stencil-based GPU programs to exploit inter-kernel data locality. The CUDA-to-CUDA transformation collectively replaces the user-written kernels by auto-generated kernels optimized for data reuse. The transformation is based on two basic operations, kernel fusion and fission, and relies on a series of automated steps: gathering metadata, generating graphs expressing dependencies and precedency constraints, searching for optimal kernel fissions/fusions, and generation of optimized code. We show how the automatic transformations were practical and effective in exploiting exposed data localities for a variety of real-world applications with large codebases that contain dozens of kernels and data arrays.

The HPCI Shared Storage is the data sharing infrastructure of HPCI. In April of 2015, we found 150000 data corrupted files in TokyoTech data servers. Even though most of files rescued by replica files, total 923 files are lost in this failure.
As the result of the analysis, this is the typical silent data corruption. To cope with the silent data corruption, the periodical data integrity check using file digest feature in Gfarm is introduced in HPCI Shared Storage. This periodical data integrity check enables data corrupted file detection within one week.

太陽は磁化した恒星であり、その表面では強磁場領域である黒点がしばしば観測される。磁場は太陽内部で乱流的なプラズマによるダイナモ運動により生成されていると考えられている。太陽黒点数には長らく謎となっている大きな謎がある。それは黒点数11年周期である。我々はまだ、このような美しい周期が何によってもたらされるか理解できていない。理解するための鍵は、太陽対流層内部にひしめく乱流的な熱対流にある。しかし、乱流は非線形で複雑な現象であるので、数値計算を用いた調査が重要となる。以前は、太陽対流層計算は大きな計算機を使う事や解像度を上げる事について問題を抱えていたが、音速抑制法と呼ばれる方法を採用する事で以前の問題を解決した。音速抑制法を用いる事で我々の太陽内部の理解は著しく向上した。講演では小さなスケールの大きなスケールの磁場を作るのにかかわる効果について説明する。

The sun is a magnetized star. We can frequently observe sunspot, i.e., strong magnetic region, on the solar surface. The magnetic field is thought to be generated in the solar interior by dynamo action of turbulent ionized plasma. The number of sunspot has a long-standing big mystery, i.e., 11-year cycle. We have not understood what makes this beautiful cycle. The key is the turbulent thermal convection in the solar convection zone. The turbulence, however, is nonlinear and complicated, and investigations by numerical calculation have important roles. Since we had problems on using large supercomputers and increasing resolution, we newly adopt a method so called, the Reduced Speed of Sound Technique (RSST) which can solve our previous problems. Using this method we have significantly improved our understanding of turbulent flow and magnetic field in the solar interior. In the talk, one example, the small-scale effect for constructing the large-scale magnetic field is shown.

東京大学大学院地球惑星科学専攻在学中に太陽対流層計算の方程式を見直す事で大規模計算機でも効率的な計算を実行可能な手法を提案した。その後、京コンピューターを利用する事で長らく謎であった太陽の表面勾配層を再現することに世界で初めて成功し、その維持機構を明らかにした。 2014年3月に博士号を取得後は学術振興会海外特別研究員としてアメリカ合衆国High Altitude Observatoryに滞在。太陽内部での磁場生成機構を研究し、これまで考えられていたよりも強い乱流的磁場が太陽内部に存在する事を明らかにした。 2015年7月からテニュアトラック特任助教として千葉大学にて引き続き、大規模計算機を用いた太陽内部の研究をおこなっている。

Hideyuki Hotta invented a method which can efficiently use large computers even for the solar convection zone with changing an equation of the problem, while he was still a graduate student at the department of Earth and Planetary Science, The University of Tokyo. Then he reproduce the near surface shear layer which was a long standing problem in the solar physics using the K computer for the first time. After he got PhD at March 2014, he stayed High Altitude Observatory, USA as a JSPS fellow. He studied generation mechanisms of magnetic field in the solar interior and revealed that there is significantly stronger magnetic field than previously expected. From July 2015, he has moved to Chiba university as a tenure track assistant professor and continues studying the solar interior with using large computers.

Precipitation forecast is important for our daily life. However, predictability of atmospheric motion has intrinsic limitations due to the chaotic nature of the dynamics. We investigate the predictability limits of convective precipitation by extracting the growing modes using the breeding method. To improve practical predictability, we developed an adaptive multi-model ensemble Kalman filter algorithm and a space-time extrapolation system with data assimilation. In this presentation, we will demonstrate the performance with synthetic and real observations.

How can I understand a program structure? How can I detect performance issues with my MPI application? Is it possible to make program refactoring easy? How do I setup development environment? How can I use existing git repository?
These are some of the topics I plan to address in my talk. I shall demonstrate how Eclipse PTP IDE can be used to speed up many aspects of parallel program development cycle. I shall show how Docker can be used for an easy setup of development environment for parallel applications that can be built and run on the K computer.
Presentation will be available at: http://www.slideshare.net/PeterBryzgalov/aics-cafe2

Asian monsoon is a planetary-scale climate system driven by differences in heat capacity between the Asian Continent and its surrounding ocean. In boreal summer, a large-scale thermal low system appears on the Asian Continent, leading to change wind direction from ocean to the continent; note that wind direction is from the continent to ocean in boreal winter. Corresponding to this change in the wind direction, rainy season begins in most region in Asia, e.g., the rainy season is known as Baiu in Japan in early summer. For improvement of the prediction of the summer precipitation in Japan, and for prediction of the precipitation change in the future climate, our research team develop a Large Eddy Simulation model (SCALE-LES). This talk introduces a series of studies related to the Asian summer monsoon. At the same time, the treatment of lower boundary condition in the SCALE-LES, which is a crucial issue of our studies, is discussed.

The K computer facilities have many features not found at other supercomputer sites. These include an expansive and pillar-free computer-room, power supply system that consists of a co-generation system (CGS) and a high-speed current-limiting circuit breaker without uninterruptible power supply (UPS), distribution boards installed not on computer-room walls but under a raised floor, extremely quiet and high-efficiency air conditioners, and water-cooling system for CPUs featuring precise temperature control. To ensure stable working of K computer and its peripherals, the facility operations and development team (FODT) of the operations and computer technologies division, RIKEN AICS is responsible for operation and enhancement of the facilities. Furthermore, FODT conducts research on the advanced management and operations of AICS’ facilities. One of the most serious problems is rapid and substantial increase in electricity prices since 2011. Therefore, we investigate the most suitable driving conditions for AICS facilities to achieve effective cost reduction. Another problem is increased power consumption by AICS. The use of electricity by AICS is strictly limited by a contract between AICS and the local electric supply company. However, recently, the facility’s power consumption exceeded the contract limit. This matter is important because the company requires us to accept a raise in the upper/lower power limit, which amounts to an increase in electricity cost. To prevent this problem, we have investigated a method to control K computer’s power consumption by using emergency job stopping together with the system operations and development team and the software development team of operations and computer technologies division, RIKEN AICS.

In order to make reliable estimations of urban disasters, such as seismic damage of structures and tsunami inundation, we need the detailed urban models for the target city. The models may include the information about all individual buildings and the underground data.
However, the quantity and quality of available single raw sources for them are limited and fragmented. Thus we need to make the detailed information pre-processing and integrating the heterogeneous raw sources, including administrative data, 3D building shapes, and a set of borehole data. The key issue is how to automate the model construction avoiding manual work.
This talk introduces some challenges in urban simulation modeling, and proposes a framework of data integration processing. The framework can automate some intermediate data conversions required in the processing.

Massively parallel computing in the field of theoretical and computational molecular science becomes more important year by year to utilize massively parallel supercomputers to perform the first-principle electronic structure calculations on large-size and complicated molecular systems. Our research team has been developed the novel molecular science software NTChem [1] for the massively parallel computing of electronic structures on the petascale supercomputer such as the K computer. NTChem involves the novel development of theory and algorithm, which make possible through the collaborative use of the K computer across the fields of computational science and computer science. In this talk, I introduce my works about the development of massively parallel algorithms and its implementations into NTChem. We present the overview of the massively parallel algorithm and its implementation of (1) the Hartree-Fock and density functional theory (DFT) self-consistent field calculations [1], (2) the time-dependent DFT excited-state calculation [1], and (3) the resolution of identity second-order Møller–Plesset perturbation (RI-MP2) calculations for energy [3, 4] and energy gradient, respectively.

参考文献:
[1] Nakajima, T.; Katouda, M.; Kamiya, M.; Nakatsuka, Y. Int. J. Quantum Chem. 2015, 115, 349–359. doi: 10.1002/qua.24860.
[2] http://molsc.riken.jp/
[3] Katouda, M.; Nakajima, T. J. Chem. Theory Comput. 2013, 9, 5373–5380. doi: 10.1021/ct400795v.
[4] Katouda, M.; Nakajima, T.; Nagase S. Proceedings of JSST 2012, 2012, 338–343.