Feasibility Study on Advanced and Efficient Latency Core-based Architecture

The University of Tokyo, Kyushu University, Fujitsu, Hitachi, and NEC in cooperation with RIKEN AICS started a two-year project funded by MEXT (Ministry of Education, Culture, Sports, Science and Technologies), Japan, where a future supercomputersystem towards an exascale supercomputer is studied.  It is assumed that the first target system will be deployed around 2018 with up to 30 MW and 2000 m2 constraints. The purpose of this project is to design a system by 2018 and reveal what kind of researches and developments must be taken into account.

 

As the first step to study feasibility of a new architecture, applications of material and climate sciences are selected. Based on the K computer architecture, a new architecture will be designed to cover a wide range of applications, and we will evaluate this using the selected applications. We will design system software stack for both the new architecture and commodity-based machines in order to assure general versatility.

Co-Design Methodology

Our study is being carried out by  intensively three cooperative cross-organizational teams; architecture design, application tuning and prediction, and system software design. The architecture design team provides its architectural parameters for the application tuning and system software teams. The application tuning and prediction team tunes the selected applications and measures performances using these architectural parameters. The system software team designs and develops the prototype of system software for both the new architecture and commodity-based machines.

The figure below shows how the application tuning and prediction team and the architecture design team collaborate. The application tuning and prediction team identifies crucial loops of the selected applications and collect performance accounting information by running them on Fujitsu FX10, a product using enhanced technologies implemented in the K computer. The target system performance is predicted using performance evaluation tools provided from the architecture design team. The architecture team redesigns the architecture based on the results. This cycle is performed every two months.