D5.5: Requirements of new user communities for the use of next generation computing systems evolving towards Exascale
Hayk Shoukourian · Zenodo (CERN European Organization for Nuclear Research) · 2018
High Performance Computing (HPC) is experiencing vast amount of changes in the road towards Exascale computing capability. These changes stretch throughout different levels: from technology and architectures to use cases. In order to attain the best performing HPC system, it is imperative that the underlying technology and architecture match the requirements of the current and emerging applications. This document aims to provide an overview of these requirements by assessing the needs of user communities and of HPC centres in terms of technologies and architectures for next generation HPC systems evolving towards Exascale. For this purpose, surveys have been conducted among recently started Centres of Excellences (CoEs) in Europe for collecting the requirements from HPC user communities. A different survey has been distributed to all PRACE Tier-0/Tier-1 HPC sites to understand how these requirements differ from the current state of the art, to determine the requirements of HPC centres, and possibly motivate related prototyping efforts. This deliverable summarizes the results of the two surveys. The most important points to note are indicated in the list below: a need for prototype systems involving heterogeneous system architectures that include new kinds of memory and parallel I/O file systems is seen by the user communities as well as by PRACE Tier-0/Tier-1 HPC centres; Graphic Processing Units (GPUs) are the most appealing accelerator systems for the user communities – a requirement which is already fulfilled by 45% of PRACE Tier-0/Tier-1 HPC sites; a shift from conventional x86 based processing technologies (which is currently dominating at PRACE HPC sites) to alternatives such are ARM, IBM Power Architecture, PTX (Parallel Thread Execution) processing technologies is foreseen for the surveyed HPC sites; containers, which are instances of an Operating System (OS) level virtualization, are getting more appealing due to their higher efficiency as compared to the full, hardware-level, virtualization; growing power density for the required heterogeneous compute nodes further motivates the need for the adoption of water cooling technologies.