Towards Convergence of Extreme Computing and Big Data Centers

Rapid growth in the use cases and demands for extreme computing and huge data processing is leading to convergence of the two infrastructures. Tokyo Tech.'s TSUBAME3.0, a 2017 addition to the highly successful TSUBAME2.5, will aim to deploy a series of innovative technologies, including ultra-efficient liquid cooling and power control, petabytes of non-volatile memory, as well as low cost Petabit-class interconnect. To address the challenges of such technology adoption, proper system architecture, software stack, and algorithm must be desgined and developed; these are being addressed by several of our ongoing research projects as well as prototypes, such as the TSUBAME-KFC/DL prototype which became #1 in the world in power efficiency on the Green500 twice in a row, the Billion-way Resiliency project that is investigating effective methods for future resilient supercomputers, as well as the Extreme Big Data (EBD) project which is looking at co-design development of convergent system stack given future extreme data and computing workloads. We are already successful in developing various algorithms and sottware substrates to manipulate big data elements directly on extreme supercomputers, such as graphs, tables (sort), trees, files, etc. and in fact became #1 in the world on the Graph 500 twice including the latest Nov. 2015 version. Our recent focus is also how to ssupport new workloads in categorizing big data represented by deep learning, and there we are collaborating with several partners such as DENSO to improve the scalability and predictability of such workloads; recent trial allowed scalablity to utilize 1146 GPUs for the entire week for a CNN workload. For TSUBAME3 and 2.5 combined we espect to increase such capabilities to over 80 Petaflops in early 2017, or 7 times faster than the K computer.