Modeling the Existing Cooling System to Learn its Behavior for Post-K Supercomputer at RIKEN R-CCS

At RIKEN, we are developing supercomputer Fugaku (formerly called as Post-K), which is the successor of K Computer, and aiming for official operation around 2021. It is expected to consume more power than K Computer, and because of the power saving functions of modern microprocessors the variation in the power consumption between the idle and peak states will be significantly larger and occur within a very short duration. However, an ordinary liquid cooling system cannot accommodate such large and rapid variations. Therefore, we are planning to employ software-based countermeasures to achieve stable operation of the cooling system. Such countermeasures will function in job management software such as schedulers, and predict the future behaviors of the water temperature with respect to scheduling information and affect the scheduling to confine the behaviors of the cooling system. Since these behaviors are sufficiently complicated that a simple estimation is unsatisfactory, we decided to employ simulations, and we model the cooling system using the popular modeling language Modelica. In this paper, we report some preliminary simulation and validation results using data from the existing cooling system of K Computer. We also discuss some of the issues of using Modelica simulations as a prediction tool.