Paper Title: AppNVM: A Framework for Application-specific FTLs on Tiered Storage

Paper Abstract: Energy consumption has become a prime concern for cloud service providers. A way to address this issue is to design systems that minimize data movement. In the context of storage systems, the idea is to avoid transferring all data to hosts, so that they can be processed as close as possible to where it is stored. Our work on near-data processing relies on a tiered architecture, where hosts are connected to open-channel SSDs via a programmable SoC-based subsystem. Open-channel SSDs allow the design of application-specific flash translation layers (FTL) thus leveraging the benefits of exposing SSD parallelism to the application. This talk focuses on AppNVM, the framework I have designed for application-specific SSD management on the SoC-based layer. The presentation is organised into four parts: (i) OX, our programmable NVMe controller; (ii) AppNVM, a flexible framework for application-specific FTLs in OX; (iii) our experimental framework based on the Dragon Fire Card; (iv) use cases, performance results and energy budgets for application-driven FTLs in a tiered storage system, compared to host-based SSD management.

Paper Author: Ivan Picoli, PhD Fellow, IT University of Copenhagen

Author Bio: Ivan L. Picoli is a PhD Fellow at IT University of Copenhagen, Denmark. Currently in his 3rd year of research and accepted as an intern at Microsoft Research Redmond. His work focuses on tiered storage systems and non-volatile memory management at rack-scale. Currently, Ivan is the creator and maintainer of OX Controller, the first open-source C-based open-channel SSD controller with near-data processing capabilities. He earned his master degree in 2015, focused on the self-tuning of Hive and Hadoop.