File System in Data-Centric Computing

TL;DR

This paper discusses the evolution of file systems in data-centric computing, emphasizing direct access to persistent memory and processing near data to improve efficiency and overcome bandwidth limitations.

Contribution

It introduces the concept of adapting file systems for data-centric computing by enabling processing near data and bypassing traditional page cache mechanisms.

Findings

Direct access to persistent memory reduces data copying overhead.

Processing near data enhances performance in data-centric computing environments.

Traditional file systems need adaptation for new memory and processing paradigms.

Abstract

The moving computation on the edge or near to data is the new trend that can break the bandwidth wall and to unleash the power of next generation NVM or SCM memory. File system is the important OS subsystem that plays the role of mediator between the user-space application and storage device. The key goal of the file system is to represent the file abstraction and to build the files' namespace. In the current paradigm the file system needs to copy the metadata and user data in the DRAM of the host with the goal to access and to modify the user data on the host side. The DAX approach doesn't change the concept but to build the way to bypass the page cache via the direct access to file's content in persistent memory. Generally speaking, for the case of data-centric computing, the file system needs to solve the opposite task not to copy data into page cache but to deliver the processing activity near data on the storage device side.