Matrix multiplication requires computer systems have huge computing capability and data throughputs as problem size is increased. In this research, an OpenCL-based matrix multiplier with task parallelism is designed and implemented by using the FPGA board DE5a-NET to improve computation throughput and energy efficiency. The matrix multiplier is based on the systolic array architecture with 10 × 16 processing elements (PEs). When data are single-precision floating-point, the proposed matrix multiplier averagely achieves about 785 GFLOPs in computation throughput and 66.75 GFLOPs/W in energy efficiency. Compared with the Intel’s OpenCL example with data parallelism on FPGA, the SGEMM routines in the Intel MKL and OpenBLAS libraries executed on a desktop with 32 GB DDR4 RAMs and an Intel i7-6800K processor running at 3.4 GHz, the proposed matrix multiplier averagely outperforms by 3.2 times, 1.3 times, and 1.6 times in omputation throughput, and by 2.9 times, 10.5 times, and 11.8 times in energy efficiency, respectively, even though the fabrication technology is 20 nm in the FPGA while it is 14 nm in the CPU.

Biodegradable plastics are attracting attention to reduce environmental impact and achieve Sustainable Development Goals, SDGs. We, the computational molecular science research team, are developing theoretical tools to estimate biodegradability of plastics to offer design guideline, by applying materials informatics. I will talk about multi-scale simulators of biodegradation. Changes of molecular weight distributions under degradation are estimated by a macroscopic Monte Carlo simulation. Hydrolysis reaction pathways of polyesters are investigated by nudged elastic band, NEB, as implemented in NTChem, a quantum chemistry package developed by us, and activation energies and heats of reaction are obtained. We are developing models to predict these values from monomer structures.

Simple model is expected to make reliable simulations, but it is not always available for social phenomena and complicated adhoc models are often used. Traffic simulation is an example. Various models are proposed and used for traffic phenomena, In this talk, traffic phenomena and their models are briefly reviewed.

The continuous improvement in processing speed in high-performance computer systems has been enabled by transistor scaling known as Moore's law. However, this trend is predicted to end in the near future. It is vital to research and develop new, more efficient high performance architectures to continue realizing high performance computing systems. One of the ways to improve performance of computer systems in the post-Moore era is utilizing domain specific architectures. In this talk, we briefly introduce our recent research efforts on a domain specific architecture for graph processing. In this research, we focus on the edge-centric graph processing model and propose a dedicated cache architecture for exploiting data locality. We have evaluated our cache architecture by a light weight cache simulation and the results showed that it can reduce the number of LLC cache misses by up to 89.9%.

In recent years, in the automotive industry, weight reductions are indispensable for complying with carbon dioxide emission regulations. Although automotive companies have been mainly using steel sheets, they want to employ multi-material structures including extrusions, castings, or 3D printings of aluminum alloy or resin to achieve weight reductions. However, the structural design will be more complex because multi-material structures have a higher degree of geometric freedom than sheet metal structures. Therefore, numerical simulations need to play a more critical role in designing optimal vehicle structures.
For the last several decades, a Lagrangian finite element method (FEM) using mainly shell formulation has been the de facto standard in the automotive industry. However, shell formulation cannot numerically model the multi-material structures mentioned above because they do not have a constant thickness. Thus, the continuum formulation has to be applied, but this approach poses two computational problems.
The first problem is that an enormous number of finite elements using continuum formulation is required to discretize the multi-material structures spatially. A scalable method in a massively parallel environment is indispensable for this simulation. Secondly, we need to spend more than a month to generate the finite element mesh of a car body. Therefore, it is challenging to investigate many patterns of vehicle structures.
Considering the background as mentioned above, we focus on a Eulerian finite volume method (FVM) [1] based on continuum formulation [2] using a scalable hierarchical Cartesian mesh method [1]. This Eulerian FVM [1] has the following three advantages. The first one is good scalability [1] in a massively parallel computing environment. Secondly, we can easily generate the computational mesh of a car body only within 10 minutes. We will demonstrate the stiffness analysis of a body-in-white structure, which is spatially discretized by approximately 200 million cells and was computed using 104,520 cores on the K computer. Thirdly, the proposed Eulerian method is easy to couple a conventional finite volume fluid solver.
In future work, we plan to conduct car crash simulations using many patterns of multi-material vehicle structures to study ultralight vehicle structures.
[1] K. Nishiguchi 2019 https://doi.org/10.1002/nme.5954 [2] K. Nishiguchi 2018 https://doi.org/10.1002/nme.5790

In this talk, I will introduce some of my previous work including my Ph.D. thesis, new OS features and middleware to make use of emerging hardware technologies. Emerging hardware devices/features (e.g., FPGA/ASICs, non-volatile memories, Intel SGX) have been widely studied due to strong demands on performance improvement, energy efficiency, and data protection of computer systems. However, due to a lack of system software support, existing computer systems cannot fully utilize these new devices. I have proposed new operating system functions and middleware, which are useful to improve the performance/usability of the devices or analyze an application behavior with them.

This presentation will introduce new optical devices that enable HPC and AI system architectures to free up from cost-prohibitive active optical cables and scale-limiting copper cables. Developed by parallel teams at HPE, VCSEL based and silicon photonics ring resonator technologies both enable passive optics at scale, and can integrate as mid-board, co-packaged or even 3D-stacked optical devices. These technologies set the course for much more capable interconnects, thanks to a significantly reduced cost structure and an energy profile tracking to sub 10 pJ/bit. Application at the system level will then be outlined, including the enablement of the Hyper-X and other multi-dimensional all to all topologies thanks to new components like fiber-shuffles.

The parallel multigrid method is expected to play an important role in large-scale scientific computing on exa-scale supercomputer systems. Previously we proposed Hierarchical Coarse Grid Aggregation (hCGA), which dramatically improved the performance of the parallel multigrid solver when the number of MPI processes was O(10^4) or more. Because hCGA can handle only two layers of parallel hierarchical levels, the computation overhead due to coarse grid solver may become significant when the number of MPI processes reaches O(10^5)- O(10^6) or more. In the present work, we propose AM-hCGA (Adaptive Multilevel hCGA) that can take into account multiple layers of three or more levels, and show preliminary results using the Oakforest-PACS (OFP) system by JCAHPC. Additionally, we also examine the impact of a lightweight multi-kernel operating system, called IHK/McKernel, for parallel multigrid solvers running on OFP.

This is a joint work with Balazs Gerofi (RIKEN R-CCS), Yutaka Ishikawa (RIKEN R-CCS), and Masashi Horikoshi (Intel).

Recently, FPGA has been attracting attention as an alternative device to accelerate HPC applications. Data flow computing model is a popular abstraction of computing in both fine-grain and coarse-grain from decades, and this model is used as a programming model for FPGA such as Maxler DFE and SPGen (Stream Processor Generator). While it is a kind of “static” data flow model, it might be interesting to extend models by using “dynamic” data flow models as old dataflow architecture to handle the dynamic behavior of systems. On other hands, global programming models to integrate FPGA computing into parallel computing of host processors are also important. OpenMP task and target directives, which are recently introduced in OpenMP 4.5, can be extended to specify the interface to offloaded computation done by FPGA. And, the optimization for FGPA needs different metric such as hardware resources, which are very different from the optimization of CPU and GPU. In this talk, issues on programming models for FPGA accelerated HPC are presented.

The K computer shut down in August with many achievements. We are now working on the installation and preparation of the operation of the Fugaku supercomputer. In the spring of next year, an early access program that uses a part of Fugaku installed in R-CCS will start. To improve user services, we are now considering "cloud service," which includes collaboration with commercial service providers. The "cloud service" is expected to contribute to improve the usability of Fugaku and increase the number of users and application areas drastically. To achieve highly energy-efficient, we optimize the Fugaku system and its facility operation based on operation data analysis and have to motivate users to increase energy efficiency by using Power API. To steadily advance these new activities, it is more critical to collaborate with research teams and operation division tightly. In this talk, I give a vision of the operation improvement towards new service on Fugaku.