Workload Optimized Systems - High Performance Computing and Analytics


High Performance Computing and Analytics


High Performance Computing

Optimization technologies for applications and system software are crucial for supercomputers, such as IBM Blue Gene/Q Supercomputer (Sequoia), which achieved 20 petaflops of peak performance in 2012 at Lawrence Livermore National Laboratory. Our research topics include application-specific optimization technologies, especially for quantum chromodynamics, fast Fourier transforms, and computational fluid dynamics.



Research on technology that solves customer’s real problems effectively and easily by using Supercomputers.

Background

  • The performance of Supercomputers is rapidly improving due to massively parallel architectures (about 1,000-fold improvement in the past ten years), and application areas for numerical simulations continue to expand.
  • The demands for software that more efficiently and more easily solves real problems is increasing as application fields expand and system complexities increase.

Our Focus

  • Application optimization technology for QCD (Quantum Chromo Dynamics), FFT, and CFD (Computational Fluid Dynamics)
  • Achieved the first place of G-FFT category at 2008 HPC Challenge Award
  • Automatic software bottleneck discovery and solution generation based on an expert‘s knowledge base
  • Optimized network protocols and hybrid simulation environments
  • Application analysis for next generation supercomputer




Analytics

We are focusing on technology to optimize performance of parallel and distributed big data processing with high throughput and low latency, such as IBM InfoSphere BigInsights to process large data using Hadoop and IBM InfoSphere to process large streaming data continuously with low latency. We also focusing on X10, a parallel and distributed programming language, to improve productivity of parallel and distributed programs.



By leveraging the analytics technologies for Big Data, we are developing scalable platform for real-time analysis of complex system behavior from large amount of events.

As base technologies, we are developing the Java backend of the parallel and distributed programming language X10, focused on the efficient implementation of X10's reified generics in Java, the distributed garbage collection for multiple JVMs, natural interoperability with existing Java programs, and mixed-mode (Java and C++) execution of X10 programs.

We are also developing a byte code compiler for the query language Jaql, which runs on top of Hadoop Map Reduce framework, and developing M3R (Main-Memory Map Reduce), which is compatible with the framework, with T. J. Watson Research Center.



Publications

  • Mikio Takeuchi et al. Compiling X10 to Java, 2011 ACM X10 Workshop, 2011.
  • Mikio Takeuchi et al. Fast Method Dispatch and Effective Use of Java Primitives for Reified Generics in Managed X10, 2012 ACM X10 Workshop, 2012.
  • Kiyokuni Kawachiya et al. Distributed Garbage Collection for Managed X10, 2012 ACM X10 Workshop, 2012.
  • Olivier Tardieu et al. X10 for Productivity and Performance at Scale - A Submission to the 2012 HPC Class II Challenge, 2012. Best Performance Award.