R-CCS Cafe 第227回 特別版

Supercomputer Fugaku has been a big success for RIKEN, Japan, and the HPC community at large. It shows that conventional wisdom, such as "All you need are GPUs", can be challenged and that we have to adapt our minds to the rapidly changing landscape of modern hardware and software. I am certain everyone who reads this abstract will know similar anecdotes from their research fields, however I will outline in this presentation some of my past and resent research which demonstrates how my colleagues and myself have challenged, and continue to challenge, the status quo in HPC interconnection networks, computer architectures, and system software.
While the successor to Fugaku is years away, it is also right around the corner. Hence, further performance studies will be required to select and fine-tune the next Japanese systems. Therefore, I will outline some of the immediate and also long-term steps all teams at R-CCS can perform to aid a structured co-design approach in the coming years. Performance "monitoring" --via instrumentation, data collection, regression testing, benchmarking, algorithmic options-- will be key to (automatically or machine-assisted) tune future HPC systems, not only in terms of peak performance but also in terms of usability and utility.