Building High-performance Storage Systems for Near-storage Accelerators

Speaker:  Jian Zhang, Rutgers University.

Time:       10:00 am, Jun. 25^th

Location: SIST1A 200

Host:        Shu Yin

Abstract:

Modern near-storage data processing technologies like computational storage devices (CSD) promise to address both hardware and software pathologies of traditional storage, such as data movement, communication, CPU bottlenecks, and others. Yet, fully exploiting the capabilities of CSD demands designing system software and firmware that can improve performance, but without compromising fundamental storage properties, like isolation, atomicity, durability, and security.  My work investigates and builds fundamental principles and techniques for near-storage data processing to accelerate data-intensive applications. In my talk, I will first present FusionFS, a high-performance filesystem for accelerating near-storage I/O and data processing. FusionFS introduces a novel storage abstraction, CISCOps, inspired by the classical CISC-based processor architecture. CISCOps combines heterogeneous I/O and data processing operations and offloads them to CSD to reduce data movement cost by 10x and improve performance by 4x compared to state-of-the-art file systems without compromising durability or resource fairness. In the second part of my talk, I will introduce OmniCache, a novel cache management solution for distributing data and processing responsibilities across the host and the near-storage accelerators. OmniCache combines the use of host and near-storage resources. Evaluation of OmniCache demonstrates significant performance gains of up to 3.24x.

Bio:

Jian Zhang is a final-year Computer Science Ph.D. student at Rutgers University. His research interests focus on storage systems, including but not limited to, file systems, near-storage accelerators, and non-volatile memories. His works have been published on top-tier system conferences, including FAST, ASPLOS, SC and SOSP