Is Storage Hierarchy Dead? Co-located Compute-Storage NVRAM-based Architectures for Data-Centric Workloads
David Roberts, Jichuan Chang, Parthasarathy Ranganathan, Trevor Mudge · 2010
The increasing gap between the speed of the processor and the time to access the data in the disk has historically been offset with deeper and larger memory hierarchies with multiple levels of SRAM, DRAM, and more recently, Flash layers for caching. However, recent trends that point to a potential slowdown of DRAM growth and the emergence of alternate resistive non-volatile memory technologies and properties of emerging data-centric workloads offer the opportunity to rethink future solutions. Specifically, in this paper, we examine an approach that leverages both the memory-like and disk-like attributes of emerging non-volatile memory technologies. We propose a new architectural building block- called nanostores- that co-locates computation with a single-level data store in a flat hierarchy, and enables large-scale distributed systems for future data-centric workloads. We present a new evaluation methodology to reason about these new architectures, including benchmarks designed to systematically study emerging data-centric workloads. Our evaluation results demonstrate significant potential for performance benefits from our approach (often orders of magnitude) with better energy efficiency.