Paper Title: HPC Driven Motivations for Ordered Key-Value Based Computational Storage

Paper Abstract: Massive data storage, management, and manipulation combined with the advent of fast network attached flash storage, combined with server compute and memory bandwidths, struggle to keep up with network and IO has pushed the exploration of computation near flash storage. Many data tasks are potentially more efficiently and economically performed when near storage devices. Scientific data — while today is mostly stored in linear bytes in files — almost always has a hidden structure that resembles records with keys and values. These data are typically persisted in an order that is not the same as retrieval, creating the need to budget time and resources for costly, massive sorting operations upon retrieval for efficient queries. Additionally, data queries in HPC almost always require ordered retrieval and frequently return many orders of magnitude less data than the full data source. We will provide multiple motivations for computation near flash storage for use in scientific computing using HPC and in particular we will describe the motivations and applications for an ordered key-value based computational storage drive collaboratively developed by SK and Los Alamos.

Paper Author: Qing Zheng, Scientist, Los Alamos National Laboratory

Author Bio: Qing Zheng is a Scientist at Los Alamos National Laboratory's High-Performance Computing Division. Qing performs basic I/O and storage research that guides the Lab's future computing platform and storage infrastructure designs. Qing obtained his PhD degree in Computer Science from Carnegie Mellon University in 2021. Qing is known for his expertise in scientific data analytics and distributed filesystem metadata management. Qing's work has been recognized by press releases, R&D 100, and Supercomputing Best Paper awards. Qing has created DeltaFS and IndexFS, and is the co-founder of Intel's Cloud Object Storage Benchmark software COSBench, now a de-facto standard when it comes to cloud object storage benchmarking. Qing's current work involves leveraging Computational Storage technologies to fast track data management and accelerate some of the world's largest scientific simulations.