国際会議SC23における展示について

情報連携推進本部および情報基盤センターでは，共同研究・共同利用拠点の提供計算機資源， サービス業務，および研究活動のアウトリーチを目的として， 2023年11月12日～17日米国コロラド州デンバーで開催される 国際会議SC23にて展示を行いました． 展示ブースでは， 13日〜15日に情報基盤センター教員および関連分野の招待研究者によるショートトークを開催しました．

・展示期間：2023年11月13日～16日
・展示ブース：♯1970　

ショートトークのスケジュール

Monday, 13 November 2023 (In Opening Gala)

• 19:15-19:30 Hiroyuki Takizawa (Tohoku University)
  • "Performance evaluation of modern vector processors with high bandwidth memory"
    Abstract: Tohoku University Cyberscience Center has just started operation of Supercomputer AOBA-1.5, which employs the 3rd generation SX-Aurora TSUBASA vector computing system as the main computing resource. This talk introduces our performance evaluation results of the modern vector computing system on SPEChpc benchmarks and practical applications. In addition, we are actively working to expand the applicable areas of the vector computing technologies, from numerical simulations to big data analysis.
• 19:30-19:45 Osni Marques (Lawrence Berkeley National Laboratory )
  • "Massively Parallel Eigensolvers Based on Minimization Strategies"
    Abstract: This presentation will summarize work in unconstrained schemes strategies for the solution of eigenvalue problems in electronic structure calculations. These schemes employ a preconditioned conjugate gradient approach that avoids an explicit reorthogonalization of the trial eigenvectors, in contrast to typical iterative eigensolvers, therefore reducing communications and becoming attractive for the solution of very large problems on massively parallel computers. The presentation will also discuss the need to rearrange calculations and FFTs to achieve performance, in particular on GPUs.
• 19:45-20:00 Takeshi Fukaya (Hokkaido University)　
  • "Numerical Investigation of Mixed Precision Sparse Solvers using low precision computing"
    Abstract: Exploiting low precision computing in the filed of numerical linear algebra is one of important tasks because the performance improvement of double-precision floating-point number (FP64) on recent hardware is difficult. In this study, for a linear system with a sparse, large, and non-symmetric coefficient matrix, we develop a mixed precision (MP) method that uses low precision computing in dominant computation and provides a solution as accurate as that by a conventional method. Our method is based on a mixed precision variant of iterative refinement (MP-IR), and we consider two candidates of Krylov method as an inner solver used in MP-IR, namely the GMRES and BiCGSTAB methods. We implement the two methods, namely MP-IR using GMRES and MP-IR using BiCGSTAB, by combining FP64 and FP32 and conduct numerical experiments on a standard CPU platform to investigate their characteristics and effectiveness. In our talk, we present the outline of our developed methods and highlights of the experimental results. This is joint work with Yingqi Zhao and Takeshi Iwashita.

Tuesday, 14 November 2023

• 13:15-13:30 Franz Franchetti (Carnegie Mellon University)
  • "Updates on SPIRAL and LibraryX"
    Abstarct: In this talk we present an update on SPIRAL and the LIbraryX framework enabled by it. We briefly discuss the X-libraries FFTX, PROTOX, GBTLX, and NTTX implemented as eDSLs (embedded domain specific languages/libraries) in C++ as well as their instantiation as plugins for LLVM/CLANG/FLANG, C++ Python and Julia. We discuss how the LibraryX framework and SPIRAL's capabilities can be utilized for developing domain-specific accelerators and the associated software stack. Finally, we discuss how LibraryX can be used in the context of generative AI to "police" systems like ChatGPT and CodePilot to ensure correctness of the implementation returned by these systems.
• 13:30-13:45 Tetsuya Hoshino (Nagoya University)
  • "Performance Evaluations of Temporal Blocking on Modern Multi-core CPUs"
    Abstract: Temporal blocking is known as an effective optimization method for stencil calculations, a calculation pattern that frequently appears in computational fluid dynamics simulations. Since the performance obtained by temporal blocking is highly dependent on the performance balance of the hardware, it is not clear what performance can be obtained on the latest multi-core CPUs with large shared caches or HBM. In this presentation, we evaluate the performance of temporal blocking techniques on A64FX and Sapphire Rapids with HBM, and Intel Xeon and AMD EPYC CPUs with large shared caches.
• 13:45-14:00 Takahiro Katagiri (RIKEN R-CCS / Nagoya University)
  • "Towards new auto-tuning technology for next-generation supercomputers: from AI to quantum-related technologies"
    Abstract: In recent years, significant strides have been made in AI technology, primarily spurred by the advent of large-scale language models (LLMs). Conversely, fundamental research into harnessing the power of quantum computers for high-performance computing platforms is rapidly gaining traction. While these two fields may appear distinct, their underlying technologies are interconnected. This relationship bears relevance, especially in the evolution of next-generation supercomputer systems. It is noteworthy that both domains inherently possess "parameters" that wield influence over performance. Moreover, the fine-tuning of these parameters is imperative for achieving peak performance. In this presentation, we aim to provide a comprehensive overview of auto-tuning (AT) technology, recognized as a pivotal factor in the forthcoming development of next-generation computers.

Wednesday, 15 November 2023

• 13:30-13:45 Yohei Miki (The University of Tokyo)
  • "Performance evaluation of N-body code on NVIDIA H100 PCIe and AMD MI210"
    Abstract: Recent competition among GPU vendors has driven GPU performance improvements. Vendors advertise the performance of their products; however, scientists in various fields should verify the actual performance in scientific simulations. We have developed gravitational N-body codes based on the direct method and have optimized for recent NVIDIA and AMD GPUs. Detailed performance evaluation reveals that (1) NVIDIA H100 PCIe is 2.0x faster than NVIDIA A100, (2) AMD MI210 is 1.4x faster than AMD MI100, and (3) both GPUs are 1.5x more power-efficient than their predecessors. We will discuss the origin of the observed speedup and the discrepancy with the ratio of the theoretical peak performance.
• 13:45-14:00 Satoshi Ohshima (Kyushu University)
  • "BLR-QR on GPU using MPS and MIG / How to utilize thousands of GPU cores?"
    Abstract: Current GPUs have massively parallel computation cores, but various applications don't have enough parallelism to fill the cores. In order to obtain high performance reasonably, we focus on the utilization of MPS and MIG features of GPU. We have confirmed the positive result on QR decomposition of Block Low Rank matrix (BLR-QR). We show the current status of the work.
• 14:00-14:15 Toru Nagai (Nagoya University)
  • "The usage environment of Azure CycleCloud and benchmark test results on virtual machines"
    Abstract: We conducted a performance measurement of virtual machines targeting Microsoft Azure with the main purpose of investigating the usage environment of public clouds. Specifically, we used Azure CycleCloud, which is specialized for HPC usage environment, under the joint research of Nagoya University Information Infrastructure Center and Microsoft Japan, and executed various benchmark programs on virtual machines. In the talk, the speaker will report on the usage environment of Azure CycleCloud and the benchmark test results on virtual machines, and discuss the collaboration between supercomputer systems and public clouds.
• 14:15-14:30 Yuki Satake (Hokkaido University)
  • "Tensor product structure preserving preconditioners for matrix equations"
    Abstract: In this talk, we consider iterative methods for solving linear matrix equations. For linear systems, there are many existing preconditioning techniques. However, most of them disrupt the tensor product structure of the coefficient matrix of the matrix equation, which increases the memory requirements. We propose memory-saving preconditioners for matrix equations by preserving the tensor product structure.

SC23展示ポスター

History of ITC-NU	Supercomputer system	HPCI and JHPCN

Feasibility Studies on Next-Generation Supercomputing Infrastructures: System Software and Library Research	Optimizations of ℋ-matrix-vector Multiplication for Modern Multi-core Processors	Usage environment of Azure CycleCloud and benchmark test results on virtual machines	Dynamic Core Binding (DCB) approach for load balancing in parallelization with MPI/OpenMP

「不老」のPV動画（英語字幕付き）