37th Workshop on Sustained Simulation Performance

Tohoku University Cyberscience Center started operation of the world’s largest SX-Aurora TSUBASA system named AOBA-S in August 2023. This talk reports the experience of operating the AOBA-S system while showing some performance evaluation results and discussions. Several important applications have already been optimized for AOBA-S, and the performance evaluation results clearly suggest the potential of the latest-generation vector engines adopted in AOBA-S. This talk also introduces our research projects that have recently started in collaboration with AOBA users.

A direct numerical simulation code for compressible turbulent flows under isothermal conditions in a box with periodic boundary conditions was developed. A finite difference method is used for a discretization of the governing equations. In paticular, an eighth-order compact difference scheme was used for the covective terms and the Mattor’s method, which is a parallel solver for a linear system with tridiagonal matrix, was used to compute the first-order derivative of the covective terms. Performance of the DNS code was measured on the SX-Aurora TSUBASA.

The Cybermedia Center at Osaka University has installed an IaaS-typed Computing Infrastructure in March 2024 and will soon start the IaaS service. In this talk, the speaker will start with the background and show the aim and strategy of this system installation from now on as well as overview the structure of the system.

The Monte Carlo (MC) method is widely used in many engineering fields, especially in uncertainty quantification, due to its robustness and simplicity. However, the existing execution pattern of MC suffers from low efficiency and scaling problems in high-performance computing (HPC). In this talk, the speaker will introduce the code intrinsic framework designed to tackle the efficiency problem of MC in HPC. The basic idea of the code intrinsic framework is to reduce the redundant calculations of MC and increase the code vectorization rate. Numerical results show that performance improvements can be achieved on various platforms, including the Intel and SX-Aurora TSUBASA machines.

Moore’s law is slowing down despite HPC centers’ increasing energy consumption. For cost and climate reasons, developing techniques for reducing energy usage while at least maintaining performance is sensible. DEPO, a software-agnostic, node-level, and dynamic power-capping approach by Krzywaniak, Czarnul, and Proficz (2022) promises to achieve these goals. We present their approach to real-world challenges based on an FEA. The Direct Tensor Computation (DTC) of Ralf Schneider and Johannes Gebert is run with DEPO to demonstrate the tools’ applicability. We explore different points in the input configuration space of the application and investigate the impact on energy consumption under power capping. Furthermore, we present and discuss different ways to continue and expand this research.

Computer Simulations that were previously regarded as CPU-bound become gradually memory-bound as the growth in memory bandwidth cannot keep up with the much higher advancements in raw computational power. This imbalance is quantified with a relative factor of approximately5.1 per decade since the 1990s, where a rise in memory bandwidth is met with a 5.1-times increase in relative computing power. In practical terms, comparing a NEC SX-4 from 1994 that operated at a balanced computational intensity of 0.125 FLOP/Byte with an Intel Ponte Vecchio accelerator from 2023 that operates at 15.9 FLOP/Byte shows a factor of 127 in the described imbalance. Mixed precision approaches that were traditionally used to speed up the throughput of calculations inside the CPU core on a cache/register level now provide speedup due to less demanded memory bandwidth. Approaches to compress arrays in a lossless or lossy manner to reduce memory bandwidth were implemented in LLNL’s zfp library, although it is not able to process the data in real-time. A similar approach targeting a real-time lossy array compression (RTLAC) algorithm is currently being evaluated for use in highly memory-bound computer simulations at HLRS.

This talk summarizes the challenges that lay ahead for HLRS in the coming decade. It looks into the challenges that we face when changing technology from CPU to GPU. It will also address the issue of AI as a new user community in HPC.

This presentation will highlight computational methodologies and results from two industrial applications related to the field of turbomachinery and steel casting. Details of the computational methods featuring, adaptive mesh refinement, dynamic load balancing and multigrid methods will be presented together with the required HPC resources.

In recent years, Ising machines have attracted much attention due to their potential in solving combinatorial optimization problems that are challenging for conventional computers. For optimization problems formulated into quadratic unconstrained binary optimization (QUBO) problems with constraints, known as constraint problems, an objective function and constraint function, along with penalty coefficients, are combined into a single Hamiltonian of a QUBO problem. However, when solving constraint problems, solution accuracy typically degrades by excessively large penalty coefficients avoiding constraint violations. As a result, minimizing the objective function becomes challenging. To solve this issue, the presentation introduces a method that partitions constraint functions and reduces penalty coefficient values. Performance evaluation using the traveling salesperson problem (TSP) with one-hot constraints illustrates that the proposed method enhances solution accuracy compared to conventional approaches.

In the last few years, Artificial Intelligence and Machine Learning have been dominant topics in the general field of computer science, and the scientific community as well. Since one of the core principles of machine learning-based algorithms requires a huge amount of data being processed on a large scale, it is expectant that high performance hardware would be needed for such computations. In this context, HPC can provide a suitable platform for various aspects of the ML pipeline, from data pre-processing to the deployment of models in production environment. In this talk, we will look at how HLRS is working towards optimizing its systems for AI and ML workloads, what are the aspects of ML pipelines that are relevant for HPC, how traditional HPC workloads, like simulations, can be integrated into ML pipelines, and what is the outlook for AI on HPC in the near future at HLRS.

The HANAMI project, a strategic alliance between Europe and Japan, aims to innovate high-performance computing (HPC) applications for next-generation supercomputers across various scientific domains. This collaboration focuses on enhancing simulation capabilities in environmental sciences, biomedicine, and materials science. Aligned with the EuroHPC Joint Undertaking, HANAMI will port existing code, evaluate application performance on new architectures, and facilitate access to advanced supercomputing resources like Fugaku and EuroHPC systems.