Source: Unsplash
While computational power is steadily increasing in modern computing systems due to the rise of multicore and accelerated systems, the relative performance of the memory subsystem is falling further and further behind. In the context of High-Performance Computing (HPC), this can be partially compensated by combining traditional main memory with a small amount of high-bandwidth memory. In systems with accelerators, heterogeneity is already visible due to the presence of different kinds of memory. However, applications must be heavily modified and use vendor-specific application programming interfaces (APIs) to exploit different kind ofs heterogeneous memory.
Thus, there is a need to develop a portable, vendor-neutral view of heterogeneous storage to enable productive use in scientific and technical applications.
To this end, the I12 / IT Center, together with the French INRIA Centre de Recherche Inria Bordeaux – Sud-Ouest, has submitted a research proposal. The project “Heuristics for Heterogeneous Memory” (H2M) is funded by the German Research Foundation (DFG) and the French funding agency Agence Nationale de la Recherche (ANR) for three years. RWTH Aachen University and the French project partner INRIA are jointly developing support for new memory technologies such as High Bandwidth Memory (HBM) and Non-Volatile Random-Access Memory (NVRAM). These technologies are increasingly being used in HPC systems as additional memory alongside the usual Dynamic Random Access Memory (DRAM).
The goal of H2M is to provide portable interfaces to identify available memories and their properties and to enable access. Based on this, abstractions and heuristics are to be developed to give application developers and runtime systems some means to control in which memory kind data should be stored and when data should be moved between different kinds of memory.
However, it is still unclear what these abstractions and heuristics should look like, and several fundamental questions need to be answered. But then, the project has only just begun – kickoff was at the end of January 2021.
The project kick-off meeting was held in videoconference.
Source: Own illustration
The research results of H2M will define a concrete development roadmap for parallel programming systems. Based on the performance research, concrete proposals will be developed at the end of the project to serve as a basis for standardization bodies’ recommendations.
For more information on the project, please visit the project website and the Agence Nationale de la Recherche website.
Responsible for the content of this article is Christian Terboven
