Research Groups
Takizawa Group
This group explores programming interfaces and their programming environment to enable an application developer to intensively optimize its time-consuming kernels while coding the other parts at high abstraction levels. Programming interfaces such as compiler directives are designed so that application developers can evolve existing HPC applications while avoiding system-specific optimizations directly applied to the application codes. If system-specific optimizations are mandatory for high performance, they are written separately from the application codes. By using the compiler directives, application codes are not rewritten from scratch but are incrementally improved. As a result, it is expected that existing software resources are smoothly migrated to new-generation HPC systems.
Takahashi Group
This group is working on the hierarchical abstraction of massively parallel heterogeneous systems by using numerical libraries. Specifically, we are researching and developing numerical libraries such as a Fast Fourier Transform (FFT) library, an Algebraic Multigrid (AMG) library and a mixed precision Basic Linear Algebra Subprograms (BLAS) library for massively parallel heterogeneous systems. When implementing the numerical libraries, we hide the architecture complexity of massively parallel heterogeneous systems with hierarchically abstraction from application developers and users, in consideration of (1) the interface, (2) the data distribution, and (3) the migration from existing libraries. Furthermore, we are also considering the fault tolerance of numerical libraries.
Suda Group
This group develops tools to support software evolution, with domain specific programming models of computational science. We are researching and developing efficient parallel algorithms for massively parallel distributed memory systems, and will provide these approaches as supporting tools using typical models, algorithms and programming techniques of computational science such as simulations and the semantics of numerical algorithms of computational science.
Egawa Group
By analyzing legacy software assets that migrate and optimize to massively parallel heterogeneous system, this group focuses on finding the reusable routines and descriptions in such programs, and categorize them as optimization techniques for frequently appeared code patterns. Specifically, the research topic is to cataloging the designs of applications and performance refactoring techniques applied to massively parallel heterogeneous systems, supporting the prompt migration of enormous software assets which are implemented so far in Japan to massively parallel heterogeneous systems.