JMLAB

10/5の R-CCS Cafe で牧野は以下のような話をすることにしました。興味がある方はどうぞ。

Title:
Toward a Science of HPC software (and hardware)

Abstract:

In the last three decades, it has been becoming more and more
difficult to achieve high, or even modest efficiency on advanced HPC
systems. As everybody here in R-CCS knows, the development of highly
efficient and highly scalable software for modern HPC systems like K
or Post-K requires multi-year effort of a team of exceptionally good
researchers. Of course, many other HPC systems are even harder to use.
Our particle simulator research team has been working on a new
approach to solve this problem through the use of framework. Our FDPS
framework allows the users to write their programs in the way very
close to the way they would write them for a single-processor machine,
and run them on large-scale parallel systems with very high
efficiency. The guiding principle for the design of FDPS is to
minimize inter-node communication and main memory access per timestep,
without increasing the floating-point operations too much. I will
discuss current state of FDPS and also our framework for stencil
calculation, Formura. If time allows, I’ll discuss how hardware design
should be changed to achieve better efficiency, not just that relative to
peak but the power efficiency which is far more important.