Performance measurement and monitoring in TSUBAME2.5 towards next generation supercomputers

Akihiro Nomura, Tokyo Institute of Technology

In order to design the architecture of next generation supercomputer, we have to know about the characteristics of applications which will run on that computer from various aspects such as computation, memory accesses, network I/O and storage I/O. Prediction of energy consumption is also an emerging topic as it is one strong limiting factor of supercomputer size. To build the performance model of those metrics, we have introduced several measurement and monitoring infrastructures including profilers and system-level monitoring system. In this presentation, we will report what we have done for this as well as what we could not measure and why.